Mihnea-Stefan David

Mihnea/Welcome

START_HERE.MIHNEA

Ideas, systems
& meaningful impact.

I'm Mihnea. I move between computer science, education, markets, and public life, guided by curiosity and the belief that difficult systems can be understood and improved.

MSDexplore · understand · build

curiosity clarity empathy

research education markets public life algorithms systems

rigorous civic-minded ambitious collaborative explorer builder

Open to conversations

Let's talk!

I am always happy to discuss interesting ideas, research, technology, education, markets, or potential collaborations.

@ Email[email protected]Best way to reach me ↗ IG Instagram@david_mihneaFollow along ↗ in LinkedInMihnea-Stefan DavidLet's connect ↗

ABOUT/MIHNEA-STEFAN-DAVID

Curious about systems.
Committed to people.

I am a Computer Science student at ETH Zürich, but the questions that interest me rarely stay within one discipline. I am drawn to the rigor of mathematics, the mysteries of physics, the power of computation, and the ways education and public institutions shape people's lives.

My story

Since childhood, I have been fascinated by the elegance of mathematics and the explanatory power of physics. Computer science became the place where that fascination could turn into something constructive: a way to move between theory and implementation, abstract ideas and real systems.

At ETH Zürich, I have explored this spectrum through theoretical computer science, algorithms, numerical methods, systems, formal methods, machine learning, and research in differential privacy. Alongside my studies, I have worked as a Teaching Assistant for Algorithms and Data Structures, Algorithms and Probability, and Numerical Methods for Computer Science. I have also supported students as a Coding Assistant and served as an Academic Facilitator for ETH's intensive PVK exam-preparation program. My exchange at the University of Pennsylvania immersed me in a different academic culture and brought new perspectives to my work.

My commitment to education extends beyond the classroom. I care deeply about educational sciences, political science, civic participation, and the institutions through which communities make decisions. From national student representation to organizing programming contests, conferences, and academic initiatives, I have learned that meaningful change requires both ideas and the patience to bring people together.

I thrive where innovation meets impact. Whether I am investigating a theoretical guarantee, debugging a program, designing a learning experience, or coordinating a public initiative, I am motivated by the same instinct: understand the system carefully, then make it better.

I am also deeply interested in economics and financial markets, especially market making. I find the field compelling because it brings together probability, algorithms, strategic decision-making, and the behavior of complex systems in real time. My next chapter at IMC in Amsterdam will let me explore that intersection from within.

MSDat the intersection of

01

Computer Science

Theory, algorithms, systems and responsible AI.

02

Education

Teaching, learning design and access to high-quality education.

03

Public Life

Political science, institutions and meaningful civic participation.

04

Markets

Economics, market making and decision-making under uncertainty.

05

Leadership

Turning shared ambitions into programs, events and communities.

06

Science & Discovery

Mathematics, physics and the curiosity to understand how the world works.

MY JOURNEY

Places that shaped
how I think.

RO

Where it started

Bucharest

Romania · Mathematics, physics, olympiads, robotics and student representation.

CH

Home base

Zürich

Switzerland · Computer science, research, teaching and intellectual independence at ETH.

US

Exchange

Philadelphia

United States · New academic culture and perspectives at the University of Pennsylvania.

NL

Next stop

Amsterdam

Netherlands · My next chapter at IMC, at the intersection of technology and financial markets.

?

The next frontier

Coming soon

The best part of a journey is not knowing exactly where curiosity will lead next.

RESEARCH/OPEN-QUESTIONS

Questions at the edge
of what we understand.

My interests span theoretical computer science and computing education. I am drawn to simple questions whose answers expose deeper tensions between information, efficiency, human reasoning and rigorous guarantees.

Current exploration · Work in progress

01 / Privacy × Mixing

Can mixing create privacy?

I am currently exploring connections between the convergence of Markov chains and differential privacy. The guiding question is whether a chain's mixing behaviour, usually controlled in total variation distance, can yield meaningful privacy guarantees expressed through stronger notions such as Rényi divergence.

mixing→stability→privacy Total variation · Rényi DP · Markov chains

RESEARCH PATH

Questions I have been following

02 / Foundations

Private PAC Learning

Understanding the sample complexity of private learners across realizable and agnostic settings, and how it is governed by combinatorial dimensions.

VC dimensionLittlestone dimensionRepDim

03 / Trade-offs

Samples vs. Computation

Investigating why information-theoretically elegant private learners can be computationally expensive, particularly for thresholds and halfspaces.

HalfspacesRuntimeLower bounds

04 / Toolkit

Geometry & Optimization

Learning from the roles of center points, convex geometry, quasi-concave optimization and private linear algebra in constructing private algorithms.

ConvexityTukey depthPrivate optimization

UPCOMING COLLABORATION University of Pennsylvania

05 / Computing Education × GenAI

Programming & Problem Solving
in the Age of Generative AI

Building on UPenn's open-ended Programming and Problem Solving course, I will investigate how generative AI changes the balance between algorithmic thinking, implementation, analysis and communication.

The goal is not simply to ask whether an AI system can outperform students. It is to understand where it contributes: generating strategies, translating ideas into code, debugging, evaluating experiments, reproducing an approach from a written report, and explaining why a solution works.

PILOT STUDY

01

Current students and historical student solutions as a human baseline

02

Claude Code working from only the problem statement and simulator

03

Claude Code receiving a small set of high-level strategic hints

04

Students using AI while documenting their prompts and independent reasoning

05

Claude Code reproducing a previous team's strategy from its detailed report

BACHELOR THESIS · COMING SOON

A new theoretical question,
still taking shape.

My bachelor thesis will be supervised by Prof. Dr. Johannes Lengler. The precise topic is still open, but it will live in theoretical computer science, with a likely focus on randomized algorithms and the mathematical ideas behind them.

Topic to be announced

ENGINEERING / PROJECTS

Problems explored
by building.

A collection of algorithmic experiments, interactive tools and software projects. Each began with a different question: how should a system communicate, compete, visualize or make a difficult idea easier to understand?

01 / PASSION PROJECT · ANDROID

Local

A privacy-first messenger where every connection begins face to face. No global search, no suggested contacts: two people become friends through an in-person cryptographic QR handshake.

KotlinJetpack ComposeEnd-to-end encryptionFirebase

01 / PASSION PROJECT · VERSION 0.1.0

LLOCAL

Stay safe.
Stay Local.

A native Android messenger built around a simple boundary: you can only connect with someone you have physically met. Local combines product design, mobile engineering and a custom end-to-end encrypted messaging system in one complete application. I designed it to feel familiar enough for everyday conversation while making its privacy model visible, understandable and difficult to bypass accidentally.

KotlinJetpack ComposeRoomFirebaseX25519AES-256-GCM

Download Android APK ↓ Current experimental build · Android 7.0+

Llocal

MS

Mihneaconnected in person

Made it home safely.21:42

Perfect. See you tomorrow!21:43 ✓✓

VOICE MESSAGE▶ ━━━━ 0:18

Write a message...↑

CONNECTION MODELOffline first.
Private by design.

01THE IDEA

A social graph that begins
in the real world.

Most messengers begin with discovery: search a username, import an address book or accept a suggestion. Local deliberately removes that layer. There is no public directory, no friend recommendation system and no way to contact a stranger from across the internet. Names are assigned locally, so the label you use for a person remains part of your own address book rather than becoming a global profile field.

Meeting is the permission.
The QR code makes it cryptographic.

02PRODUCT

MESSAGE

Rich conversations

Text, cropped or rotated photos, files up to 10 MB and recorded voice messages. Replies carry an encrypted copy of the quoted context, while six reactions, editing and deletion are synchronized as encrypted control events.

CONTROL

Ephemeral by choice

Conversation timers range from one minute to one week or can remain disabled. Expiry is represented in the data model rather than merely hiding a bubble, and deletion-for-everyone propagates through encrypted tombstones.

CONTEXT

Useful, not bare

Search inside a conversation, swipe to reply, double-tap for a quick reaction, inspect three delivery states and open images in a full-screen viewer with pinch and pan. Privacy does not need to mean an unfinished interface.

RESILIENCE

Sends that survive navigation

Media uploads continue outside the chat screen through an application-owned coroutine scope. The interface shows optimistic progress, applies a 20-second timeout and keeps enough state to offer a meaningful retry rather than silently losing work.

02BPRODUCT PRINCIPLES

Privacy expressed through product decisions.

The strongest privacy mechanism is not useful if normal people cannot understand when it applies. Local therefore treats trust, expiry and device protection as visible parts of the experience rather than settings hidden behind technical language.

01No invisible discovery

The application never uploads an address book to construct a graph of possible contacts. A friendship exists only after the two-device ceremony completes.

02Local meaning stays local

Contact names, resolved notification names and cached conversation context are owned by the device. The backend routes encrypted objects; it does not become the canonical social profile.

03Security has understandable states

Pending invitations expire after one hour, delivery has explicit stages, maintenance is communicated directly and destructive actions propagate rather than pretending to be complete locally.

03TRUST CEREMONY

Two scans. Two ephemeral keys.
One shared relationship.

Adding a friend is a short in-person protocol rather than a lookup. A fresh identity matrix and ephemeral X25519 key material bind the two devices to the same physical encounter. The invitation can only be confirmed by the intended inviter, must match the server-held matrix hash and expires after one hour. Scanning your own code or presenting an invalid confirmation is rejected.

DEVICE AMSINVITEephemeral key A

PUBLIC IDJwTNp5s1

JwTNp5s1 veyQIPMu IuzmbwXk spKHbOoo iEDMttMJ 6USFzmJ0 sYPDBuOY z7NAtbr6

56 random characters regenerate with every invite.

DEVICE BPCONFIRMephemeral key B

SCAN

VALIDATE

DERIVE

01Invite

Device A generates a fresh 8 × 8 matrix, public ID and ephemeral X25519 public key.

→

02Scan

Device B scans locally, chooses a private contact name and joins the one-hour session.

→

03Confirm

Device B displays the matching session and its own ephemeral public key for the return scan.

→

04◉—◉Derive

Both devices validate the matrix hash and independently derive the same initial root key.

INITIAL ROOTDH = X25519(ephemeralPrivate, peerPublic)rootKey = HKDF-SHA256(matrix, DH, "Local-Init-RK")The server stores the matrix hash for validation, never the matrix itself.

04ENCRYPTION

A new message key
for every message.

Local implements a custom Double Ratchet design in the Signal protocol family. Each message advances a symmetric chain; a change in speaking direction triggers a new X25519 Diffie-Hellman ratchet. AES-256-GCM supplies authenticated encryption and every nonce is generated independently.

The encrypted wire object carries ciphertext, nonce, authentication tag, current public ratchet key, message number and previous chain length. Skipped keys support up to 1,000 out-of-order messages. Ratchet transitions are protected by a per-friendship mutex so decryption and persistence remain atomic even when network events arrive close together.

ROOT KEYRK_n

HKDF

CHAIN KEYCK_nHMAC-SHA256

MESSAGE KEYMK_n

NEXT CHAINCK_n+1

AUTHENTICATED CIPHERAES-256-GCM12-byte nonce · 128-bit tag

04BMESSAGE JOURNEY

From tap to trusted local state.

A message crosses several ownership boundaries. The implementation keeps plaintext, cryptographic state, network transport and interface feedback separate, then reconnects them through explicit repository operations.

01Compose

The chat creates an optimistic local item immediately, keeping interaction responsive.

→

02Ratchet

The send chain advances and produces a one-time message key.

→

03Encrypt

AES-GCM returns ciphertext plus integrity data; plaintext never enters Firestore.

→

04Transport

Firestore stores the encrypted envelope and FCM sends a content-free wake-up signal.

→

05Receive

The peer serializes ratchet access, decrypts, validates and persists the new state.

→

06Clean up

After successful decryption, the transient server copy is deleted.

05MEDIA

Large files use envelope encryption.

Photos, documents and voice recordings receive a random 32-byte data-encryption key. The encrypted blob goes to cloud storage; only a compact envelope containing the key, nonce and authentication tag travels inside the ratcheted message.

DEVICEOriginal mediaphoto · file · audio

+

RANDOM32-byte DEKone key per blob

→

STORAGEEncrypted bytesopaque UUID path

MESSAGE ENVELOPEDEK + nonce + tagencrypted by Double Ratchet

06SYSTEM

Native, local-first architecture.

A single-activity Compose interface sits above screen-specific ViewModels and long-lived repositories created by the application container. Room is the source of local continuity; Firebase provides passwordless email-link authentication, encrypted transport, blob storage and content-free push signals. The result is a client that remains coherent across navigation, background transitions and process-level lifecycle events.

INTERFACEJetpack Compose

Material 3 · Navigation · CameraX · ML Kit · Coil

↓

APPLICATIONMVVM + repositories

Auth · friends · chat · media · ratchet · lock · session

↓

ON DEVICERoom v9

Messages, friendships, cached plaintext and wrapped ratchet state

TRANSPORTFirebase

Auth, Firestore ciphertext, Storage blobs, FCM and one Cloud Function

07IMPLEMENTATION

01

Atomic ratchet state

A mutex serializes cryptographic work per friendship. A received message is decrypted and its advanced ratchet state is persisted as one logical operation.

02

Granular recomposition

Per-message plaintext, quote and reaction state lives in Compose snapshot maps, allowing one changed bubble to update without rebuilding the conversation.

03

Fast chat opening

Cached plaintext is batch-loaded from Room when a conversation opens, then synchronized with incoming encrypted events from Firestore.

04

Durable asynchronous work

An application-level coroutine scope owns important sends and uploads, so leaving a screen does not silently cancel user work.

05

Explicit data evolution

The Room schema is versioned explicitly and destructive fallback is disabled. Database changes must be represented as deliberate migrations.

06

Release-oriented build

The Android release enables R8 minification and resource shrinking; a signed release APK is already produced from the current codebase targeting Android 14 while supporting Android 7.0 and newer.

07

Controlled session ownership

A session repository coordinates passwordless authentication and a single-device policy. Logging in elsewhere invalidates the previous device and removes the local state it should no longer retain.

08

Typed control events

Reactions, timer changes, edits and deletions reuse the encrypted message channel as explicit event types, keeping conversation state synchronized without introducing a plaintext side channel.

09

Operational safeguards

A live remote configuration supports normal, maintenance and mandatory-update states, with a cached last-known value so critical release decisions do not depend on one screen load.

08DEVICE DEFENCE

Privacy continues after decryption.

End-to-end encryption protects data in transit. Local also adds controls for the moment when decrypted information is present on the phone.

SCREENCapture protection

FLAG_SECURE blocks screenshots, screen recording and sensitive previews in recent apps.

ACCESSPIN + biometrics

An optional six-digit PIN, biometric unlock and automatic lock after 30 seconds in the background.

EMERGENCYDuress PIN

A separate panic code can silently erase conversations before opening the application.

SESSIONOne active device

A login on a new device invalidates the previous session and clears its locally held application state.

KEY STORAGEHardware-backed wrapping

Ratchet roots, chain keys, private DH material, matrices and skipped keys are wrapped with Android Keystore AES-256-GCM.

CREDENTIALHardened PIN storage

PBKDF2-HMAC-SHA256 with 120,000 iterations, a random salt and encrypted preferences protects local PIN verification.

09LIFECYCLE

Privacy also means deleting correctly.

Account removal, friendship deletion, expiry and maintenance states are modeled as system behavior rather than interface-only actions.

TTL

Undelivered text expires after 30 days and media after five; a successfully decrypted Firestore copy is removed immediately.

TOMBSTONES

Friendship and account deletion propagate to other devices, triggering local removal of messages, media and key material.

KILL SWITCH

A live remote configuration can place the application in normal, maintenance or mandatory-update mode.

NOTIFICATIONS

Push payloads contain no message text. Sender identity is resolved from the local database and notifications are coalesced.

10THREAT MODEL

THE SERVER DOES NOT RECEIVE

Conversation content

Message plaintext, media plaintext, contact names, the QR identity matrix or unwrapped cryptographic keys.

THE SERVER CAN OBSERVE

Routing metadata

Participants, sender identifiers, timestamps, message type, ciphertext size, ratchet headers and storage paths needed to deliver data.

11ENGINEERING STATUS

Built seriously.
Described honestly.

Local is a substantial working prototype, not yet a security-audited product. Its custom cryptographic protocol must receive independent review, automated tests and wider real-device validation before it should be trusted in high-risk environments.

IMPLEMENTEDComplete messaging flow

IMPLEMENTEDSigned Android release

NEXTIndependent security audit

NEXTAutomated test suite

NEXTCross-vendor device validation

02–05 / COURSE SERIES

Programming &
Problem Solving

Four open-ended challenges combining strategy, implementation, experimentation and clear communication.

University of Pennsylvania

02 · ALGORITHMIC CHALLENGE

Mosquito

Searching for coordinated light layouts inside a stochastic physical simulation.

?

03 · ALGORITHMIC CHALLENGE

Spy

Reasoning with partial information and adapting a strategy as evidence changes.

04 · ALGORITHMIC CHALLENGE

Football

Turning a competitive setting into algorithms, experiments and measurable decisions.

05 · ALGORITHMIC CHALLENGE

Organisms

Exploring emergent behavior through simulation, strategic design and iteration.

02 / CASE STUDY · SPRING 2026

Guiding
Mosquitoes

A stochastic optimization challenge about placing timed lights and a collector, explored through structured search, simulation and eight generations of a Java player.

JavaC++Combinatorial searchSimulation

View code on GitHub ↗

20mRELAY EDGE

01THE PROBLEM

Corral a stochastic population with a handful of lights.

Mosquitoes move independently through a 100×100 continuous world containing walls. Without a visible light they perform a random walk; when an active light is within 20 metres and unobstructed, their next one-metre step is biased toward the nearest such light by a random angle of up to 30 degrees.

The player chooses one collector and up to L point lights. Every light has a position and a periodic schedule described by its cycle length, on-time and start phase. The simulation ends when at least half of the mosquitoes enter the collector's tiny radius-0.5 capture disk.

The objective is to minimize expected rounds-to-half across random spawn positions and movement noise, while returning a legal configuration within a strict single-CPU time budget.

100 × 100A continuous board whose walls block movement and line of sight.

20m radiusOnly the nearest visible active light influences a mosquito.

±30° noiseAttraction creates drift rather than deterministic motion.

50% targetThe score is the time required to capture half the population.

COURSE MATERIAL

The complete simulation rules.

Read the original specification for motion, visibility, obstacles, light scheduling, collector placement and scoring.

SYSTEM VIEW

One lattice, many formations

Eight directional links form the common vocabulary behind every formation. Selecting different subsets of the same local neighbourhood lets the swarm assemble lines, corners, clusters and branching structures without a central controller.

LineCornerClusterBranch

02WHY IT IS HARD

Coupled decisions

A light's value depends on every other light, wall occlusion and the collector. Useful relay chains may contain components that perform poorly alone.

Noisy evaluation

Identical layouts vary across seeds because both spawning and every movement step are random. One lucky simulation can rank a worse candidate first.

Finite search time

Every additional trial consumes the same wall-clock budget needed to generate, refine and validate alternative configurations.

Lights are not isolated attractors. They are a relay system rooted at the collector.

PHYSICAL INSIGHT

An always-on terminal light secures the capture region. Helper lights extend its basin outward and alternate in time, handing mosquitoes inward along branches.

03MD1 → MD8

Eight iterations, each removing a mistaken assumption.

The final system emerged by measuring failures and simplifying the representation, timing model and budget policy rather than by adding layers indiscriminately.

MD1Named families

Enumerated lines, clusters, rings, arcs and funnels on a coarse anchor grid.

MD2Random sampling

Replaced biased nested-loop enumeration with broader deduplicated sampling.

MD3Relay physics

Stopped searching timing and committed to a terminal light plus inward waves.

MD4Lattice trees

Unified lines, clusters, bends and branches as trees on one 8-direction graph.

MD5Forward growth

Removed backtracking and zigzags with a forward-half-plane constraint.

MD6Branch timing

Let short and long branches pump independently at their natural cadence.

MD7Parity schedule

Reduced timing to two phases based on branch-depth parity.

MD8Adaptive budget

Measured evaluation cost and sized the search to each map at runtime.

04SEARCH REPRESENTATION

Discretize first.
Recover continuity later.

Instead of optimizing real-valued coordinates and independent schedules directly, MD8 samples rooted trees on a reusable lattice graph. The collector is the root, each node is a light, and parent pointers encode the relay topology.

CANDIDATE rooted tree ⊂ 8-direction lattice

20 edge length6 maximum depth8 directions

After the discrete sweep identifies promising regions and shapes, helper coordinates are perturbed locally in continuous space and re-evaluated.

05JAVA SEARCH PIPELINE

Spend precision only where it matters.

A multi-stage evaluator keeps broad exploration affordable and reserves expensive multi-seed evidence for the strongest candidates.

BuildConstruct feasible lattice sites and relay edges around walls.

MeasureTime one fallback evaluation and derive the affordable candidate count.

SampleGrow distinct forward-only trees across collector anchors.

CoarseEvaluate every candidate cheaply with one simulation seed.

RefinePerturb the best trees off-lattice into nearby continuous layouts.

ValidateRe-evaluate finalists across multiple seeds and commit the winner.

06COORDINATED TIMING

A graph coloring becomes a control policy.

With 20-metre edges, lights at equal branch-depth parity are at least 40 metres apart along a chain. They can therefore share a phase without locally competing for the same mosquito.

PHASE Adepth 1depth 3depth 5

⇄

PHASE Bdepth 2depth 4depth 6

The root light beside the collector remains permanently on, while helper levels alternate on a fixed 78-round cycle.

07SYSTEMS ENGINEERING

When search quality became a throughput problem.

After MD8 stabilized, the remaining lever was the cost of each simulation. I ported the simulator core to C++ and connected it to the Java player through a persistent synchronous subprocess.

The port mirrors important Java semantics, including the 48-bit java.util.Random generator, movement conventions, strict collector containment and collision checks. A text protocol keeps the bridge inspectable, and every failed C++ request falls back safely to Java.

JAVAPlayer_MD8_CPPcandidate search and backend selection

stdin
⇅
stdout

C++mosquito_simlong-running simulation server

08MY CONTRIBUTION

Owning the core solution end to end.

I designed and implemented the core approach, including all eight player generations from MD1 through MD8, the lattice-tree representation, relay policies, search and refinement pipeline, adaptive budget and C++ simulator.

I led the integration of the components and coordinated and structured most of the final report. The project combined algorithm design with substantial software engineering: modelling, iterative experimentation, Java implementation, cross-language IPC, behavioural validation and graceful fallback.

09TOURNAMENT RESULTS

Strong ideas, uneven robustness.

The final tournament exposed both the value and the limits of the approach. Relay trees were excellent on open or familiar maps with small-to-moderate light budgets, but the search became fragile on wall-heavy worlds, especially when many branches could be occluded.

We finished 8th of 9 overall. Rather than hiding that result, it provided the clearest diagnosis: complex maps made each evaluation slower exactly when the search needed a larger, not smaller, candidate pool.

2First-place finishes across 55 tournament configurations.

65.2Winning score on Blank with 10 lights, ahead of 72.3.

#1Mosquito map with 3 lights: 4,680 rounds.

7.2Average rank, revealing weak robustness on complex maps.

10WHAT WE LEARNED

Representation changes the problem

A carefully constructed discrete space can make an apparently intractable continuous search practical.

Measure the evaluator

Optimization quality depends on systems performance. The cost model and fallback behavior are part of the algorithm.

Test the hard regime

A strategy that dominates clean boards may fail structurally when obstacles, occlusion and larger light counts interact.

04 / CASE STUDY · SPRING 2026

Parallel
Football

A competitive multi-agent programming challenge about coordinating a team of autonomous players on a shared board, implemented and evaluated in Java.

JavaMulti-agent systemsHeuristic optimization

View code on GitHub ↗

32 × 32SHARED GRID

01THE PROBLEM

A race for 1,020 footballs.

Parallel Football is played by four teams on a common 32×32 grid. Each team owns one corner, called its home cell, and controls P players. Every non-home cell initially contains one football, creating 1,020 shared footballs that all four teams are trying to score.

At each turn, every player chooses exactly one action: move by one cell in any of the eight directions, remain still, or kick one football from its current cell to any target within Euclidean distance K. A ball reaching the team's home cell scores immediately and is removed from the game.

The important detail is that all teams submit their orders using the same observed board state, and the simulator processes those orders simultaneously. A target that looked available may therefore be moved by an opponent before our player arrives.

32 × 32One shared grid with a home cell in each corner.

P playersThe tournament changes the number of agents controlled by each team.

K radiusKicks use Euclidean distance; movement allows one diagonal or orthogonal step.

One actionEach player either moves, waits or kicks one ball per turn.

Full stateOur controller sees all players, football locations and scores before deciding.

Shared cellsPlayers and footballs may stack, creating congestion and contested actions.

COURSE MATERIAL / 1 PAGE

Read the complete original statement.

The full specification contains the exact movement and kicking rules, simultaneous-action semantics, tournament setup and the questions that guided the project.

LIVE SYSTEM

Coordinate a moving field

Players continuously collect, relay and redirect contested footballs toward their home corner. The animation captures the core challenge: many simultaneous local actions must combine into useful global flow while three opponents interfere with the same resources.

our teamopponentsfootball

HOME OUR FLOWcollect → relay → score

02WHY IT IS HARD

The rules are local.
The consequences are global.

A good controller must clear nearby balls efficiently while maintaining coverage of distant regions. It must decide whether a dense stack is worth pursuing, whether an opponent will arrive first, and whether moving a ball closer to home actually leaves a teammate available to finish the job.

01

Contention

Four teams compete for exactly the same balls, so stale targets waste turns.

02

Coordination

Sending teammates to the same cell can duplicate work and create congestion.

03

Different regimes

Small P or K rewards local racing; large P and K make assignment and relay behavior more important.

03HOW WE SOLVED IT

Start simple.
Measure. Adapt.

We implemented several Java controllers, watched them play repeated tournaments and refined the decision rules based on concrete failure modes. The final submission, HybridPlayer3, switches between two complementary strategies instead of forcing one policy to handle every configuration.

01

BasicPlayerMD

Our first complete baseline distributed players across quadrants, assigned distinct nearby balls and kicked directly toward home.

02

GreedyPlayer3

A richer local score incorporated ball density, distance to home, incoming teammate kicks and opponent congestion.

03

TwiceImprovedPlayer

A global value-based strategy scored every player–ball pair and greedily selected distinct high-value assignments.

04

HybridPlayer3

Use GreedyPlayer3 when P≤10 or K≤2; otherwise use TwiceImprovedPlayer for crowded, longer-range games.

“If the board is sparse or movement is restricted, just race. If the board is crowded and kicks matter, think strategically.”

CORE INSIGHT

No single policy dominated every configuration. Robust performance came from recognizing which version of the game was currently being played.

04JAVA IMPLEMENTATION

From game state to a Java move array.

Each strategy is implemented as a Java player controller on top of the course simulator. On every turn it reads the board, computes one instruction for each teammate and returns the complete set of actions before any team sees the others' choices.

Language Java Input Complete simulator state Output P simultaneous actions

ObserveCache football counts, scores and all teammate and opponent positions.

→

ValueCompute movement distance, kicks-to-score, density and competition for each target.

→

AssignSort candidate player–ball pairs and select distinct assignments greedily.

→

ActMove one step, score directly, or choose a relay-friendly intermediate kick.

05SCORING MODEL

Giving every target an interpretable value.

For global assignment, we estimated the effort needed to reach and score a ball cell. Player movement uses Chebyshev distance because diagonal movement costs one turn, while kick feasibility remains Euclidean because the simulator defines a circular kick radius.

PLAYER–BALL SCORE value ÷ (travel + λ(P) · kicks + 1)

Ball stacks and cells in our quadrant increase value. Long kick chains become more expensive as P grows, and targets already favored by a closer opponent receive a competition penalty.

06GLOBAL ASSIGNMENT

Choose opportunities, not simply nearby balls.

TwiceImprovedPlayer first scans the board and builds a feature record for every cell containing footballs. It estimates travel time, kicks-to-score, stack size, territorial advantage and the closest opposing player.

It then scores every feasible player–ball pair, sorts the pairs globally and greedily selects assignments while enforcing one target per player and one player per target. This avoids the duplication produced by independent local decisions without paying for an exact matching algorithm on every turn.

VALUEB(c) · q(c)

Stacks are worth more, with a mild bonus for balls in our quadrant.

TRAVELd∞(p,c)

Chebyshev distance matches one-step diagonal player movement.

CHAIN RISKλ(P) · kicks

Long kick chains become increasingly expensive in crowded regimes.

COMPETITIONclosest opponent

Contested low-value targets are discarded before they waste movement.

07KICK PLANNING

Where the ball lands matters as much as how far it travels.

If home is inside the Euclidean kick radius, the action is immediate. Otherwise, the controller evaluates every legal landing cell and balances progress toward home, proximity to teammates, distance from opponents and use of the available range.

This creates implicit relays without reserving rigid bucket-brigade roles. A teammate may continue a promising ball naturally, while the strategy remains robust when simultaneous opponent actions invalidate a planned hand-off.

01 maximize progress02 preserve relay options 03 avoid immediate steals04 use kick range efficiently

08ITERATION & TESTING

The simulator exposed what equations did not.

We repeatedly ran local and head-to-head tournaments, watched complete games and changed one heuristic at a time. This process revealed failures that were difficult to anticipate from the rules alone.

The moat

Players moved outward after kicking balls near home, leaving a ring of almost-scored balls unattended. Explicit home proximity and scoreability bonuses fixed it.

Stale races

Targets occupied by opponents often disappeared before our player arrived. Congestion and competition penalties reduced these wasted trips.

Hidden density

Binary “ball or no ball” logic ignored valuable stacks. Using diminishing returns on ball count captured density without overwhelming distance.

09MY CONTRIBUTION

Building the baseline that shaped the project.

I designed and implemented BasicPlayerMD, the team's first complete controller. It introduced quadrant-based initial placement, greedy ball selection, distance-aware kicks and lightweight coordination constraints that prevented teammates from chasing the same ball or moving into the same destination.

I also implemented and explored RelayFirst, GreedyAssist and GreedyAssistLee, testing whether more explicit passing behavior could improve long-distance scoring. These experiments showed that plausible teamwork ideas can be fragile when every team moves simultaneously.

The baseline's strong and interpretable performance helped establish the direction of our later strategies. Other teams used related greedy ideas as benchmarks, subroutines or targets for empirical tuning.

10EVIDENCE

Strong because it was consistent.

The final hybrid was evaluated across ten combinations of player count P and kick radius K. It remained near the top across sparse races, crowded boards and long-range kicking regimes.

9/10Top-two tournament finishes

+17%GreedyPlayer3 over the original GreedyPlayer in head-to-head testing

398.9Average score in the strongest local-race configuration, P=46 and K=1

11WHAT WE LEARNED

Robustness beats elegance

In adversarial simultaneous environments, a stable heuristic can outperform a sophisticated plan that depends on fragile assumptions.

Watch the system

Several of the most important improvements came from observing matches, including fixing the unattended “moat” of balls around home.

Know the failure mode

The extreme P=255, K=32 regime exposed the limits of greedy assignment and points toward explicit area control or learned policies.

03 / CASE STUDY · SPRING 2026

Soldier, Soldier,
Soldier, Spy

A Java multi-agent system for exploration, trustworthy information sharing and consensus in a partially observable world containing an unknown adversary.

JavaDistributed systemsAdversarial reasoning

View code on GitHub ↗

PT ???

100 × 100PARTIAL KNOWLEDGE

01THE PROBLEM

Build a shared truth without knowing whom to trust.

A commando team is dropped onto an unknown 100×100 grid. The soldiers must discover a hidden package and target, reconstruct the terrain and agree on a path that uses only good cells. Water is impassable and muddy ground may be crossed by soldiers, but never while carrying the package.

Each player sees only a radius-three neighborhood. Global knowledge exists only through encounters, where agents exchange records containing the cell, its reported state and the complete chain of players and timestamps through which that claim travelled.

One player is secretly a spy. It uses the same interfaces as everyone else, but may fabricate terrain, invent objectives and support dangerous paths. Three soldiers must still meet at the package and agree on exactly one valid route before the timeout.

Local visionEvery agent observes only a small region and builds its own evolving map.

ProvenanceEvery shared claim carries its origin and complete communication history.

One adversaryThe spy can lie selectively, remain plausible and exploit uncertainty.

Three votesA unique path needs support from three players gathered at the package.

COURSE MATERIAL

Read the complete rules.

The original statement defines movement costs, visibility, record provenance, communication, path voting, scoring, failure conditions and the spy's capabilities.

MEMORY MODEL

A map that remembers uncertainty

Each board coordinate stores more than a guessed value. The agent tracks provenance, confidence and contradictory reports, while separate indexed structures maintain player suspicion and mission-level state. That makes every move explainable and revisable.

CellKnowledge[100][100] O(1) coordinate lookup

→

SELECTED CELL RECORD

terrainValueenum

pointValueint

confidencedouble

directSeenboolean

conflictsSet<Value>

supportMap<Value, Set<Id>>

GLOBAL INDEXESsuspicion[playerId]packagePositiontargetPositionvisitedCells

02WHY IT IS HARD

Partial observability

No player sees the global state. Exploration must turn fragmented observations into a useful map before time runs out.

Adversarial information

The attacker manipulates beliefs rather than the grid, so plausible misinformation can be more damaging than obvious lies.

Distributed consensus

Strict validation protects the team, but excessive caution can deadlock the vote and fail just as decisively.

03SOFTWARE ARCHITECTURE

A modular Java decision system.

The controller separates knowledge management from movement, planning, communication and voting. The same state is consumed by distinct soldier and spy policies, allowing each subsystem to be inspected and refined without rewriting the entire player.

State 10,000 CellKnowledge objectsInput Observations, records, proposals and vote resultsOutput Movement, messages, paths and votes

01Knowledge

Store terrain, point type, confidence, source sets, conflicts and direct observation.

02Communication

Fuse incoming records and share only sufficiently reliable, conflict-free information.

03Navigation

Explore under uncertainty, then use cost-aware shortest paths to reach the package.

04Consensus

Validate proposals and adapt voting as the system approaches deadlock or timeout.

04KNOWLEDGE & TRUST

Evidence is not just a value.
It has a history.

Each cell tracks terrain and point type independently. Direct observations become immutable local ground truth; indirect reports contribute only through their original source. Repeated copies of the same claim therefore cannot manufacture confidence.

CELL (x, y)GOODconfidence 2 / 3conflict: false

source 4→player 2→meONE independent origin

Direct observation wins.

Contradictions against a personally observed cell increase the originator's suspicion score. Sources crossing the threshold are ignored in future updates.

05PROBLEM-SOLVING PIPELINE

Explore until the evidence supports commitment.

The player continuously switches between gathering missing knowledge and acting on a safe solution.

ObserveWrite direct terrain and objective information into the local map.

→

IntegrateAggregate distinct origins, retain ties as conflicts and update suspicion.

→

ExploreUse cost-aware search to reach the nearest insufficiently known cell.

→

PlanRun Dijkstra over high-confidence GOOD cells from package to target.

→

CommitMove to the package, propose the path and enter distributed voting.

06CONSENSUS UNDER PRESSURE

Correctness first.
Progress before timeout.

A single fixed voting rule was too brittle. We designed a phase-based strategy that starts conservatively, narrows competing proposals and eventually relaxes its requirements when waiting becomes the larger risk.

NORMALStrict

Support only fully validated paths, then progressively converge toward the best trusted leader.

→

STALLPopular

Choose the most-supported path that passes relaxed endpoint, adjacency and direct-safety checks.

→

DESPERATEDecisive

Always select a minimally valid candidate, preventing indefinite deadlock near the timeout.

07THINKING LIKE THE ADVERSARY

Deception tailored to the recipient.

The spy does not broadcast random noise. It estimates what each player already knows and constructs a separate fake package, fake target and plausible GOOD corridor for every recipient, avoiding cells likely to contradict direct observations.

01

Personalized worlds

Per-recipient campaign state makes each fabricated scenario internally consistent.

02

Forged provenance

Fake originators make misinformation appear independently corroborated across meetings.

03

Truthful cover

A small set of honest local records makes the overall message look less anomalous.

04

Adaptive sabotage

If a fake path loses support, the spy inflates a real valid path with loops, worsening the score without obvious invalidity.

08ENGINEERING THE SYSTEM

Debugging beliefs, not only outputs.

Traditional unit tests were useful for pathfinding, conflict resolution and suspicion updates, but system behavior depended on team composition, information propagation and an adaptive attacker. We therefore complemented component tests with repeated simulations and tools that exposed the internal state.

Knowledge visualizer

A browser-based 100×100 map showed confidence, conflicts, exploration frontiers and the full provenance chain for a selected cell. It revealed subtle source-tracking bugs hidden by aggregate scores.

Diagnostic maps

Three custom environments created divergent exploration paths, isolated corridors and bottlenecks to stress-test trust and consensus under genuine information asymmetry.

Empirical evaluation

Clone tests, soldier-with-spy tests and mixed-cohort round robins separated cooperative quality from adversarial disruption across seven maps.

09MY CONTRIBUTION

Designing the foundation and its adversarial counterpart.

I designed and developed G7_MD, the initial player that became the foundation of the final system and established its overall architecture.

I implemented the cell-level knowledge model, independent-source tracking, conflict-aware updates, communication protocol and the exploration strategy retained by the final player.

For the spy, I helped move fake objectives from one global story to personalized per-recipient deception, implemented loop-based path inflation and designed the fake-originator protocol. I also created the three custom evaluation maps and wrote substantial parts of the problem formulation, design and implementation sections of the report.

10RESULTS

Strengths that depend on the information structure.

The honest player performed best on structured maps where reliable knowledge accumulated quickly. The spy was strongest when uncertainty remained high enough for personalized false worlds to survive scrutiny.

#1Soldier rank on the structured circle map.

+2,100Highest spy score delta on the kr map.

5/7Maps with above-average spy disruption percentile.

60%Mean percentile for the spy across tested maps.

11WHAT WE LEARNED

Provenance matters

Message count is not evidence count. Tracking original sources prevents repeated propagation from creating false certainty.

Deadlock is a failure mode

A perfectly cautious agent can still lose. Robust systems need an explicit policy for trading certainty against remaining time.

Observability changes testing

In distributed software, inspecting internal beliefs and communication histories can be more informative than checking final outputs alone.

05 / CASE STUDY · SPRING 2026

Adaptive
Organisms

A Java multi-agent simulation about survival, learning and population control in an unknown, changing world.

JavaAdaptive agentsDistributed learning

View code on GitHub ↗

LOCAL VIEWGLOBAL BEHAVIOR

01THE WORLD

Survive without seeing the world.

Each organism lives on an unknown m×n toroidal grid and runs its own instance of a Java “brain.” It gains energy by eating food and spends energy by waiting, moving or reproducing. When its energy reaches zero, it dies.

An organism sees only the four adjacent cells, the food beneath it, its current energy and a subset of simulator parameters. It does not know the grid dimensions or the probabilities p and q governing food appearance and growth.

The goal is not merely rapid reproduction. A population that expands too aggressively may consume its resources and collapse; one that behaves too cautiously may lose territory and food to competing species.

MoveTravel one orthogonal cell on a world that wraps around its edges.

WaitConserve position while paying the environment's staying cost.

ReproduceSplit energy with a new organism placed on an adjacent free cell.

CommunicateExpose an 8-bit external state visible to neighboring organisms.

PDFOrganismsORIGINAL SPECIFICATION · 2 PAGES

COURSE MATERIAL

The complete simulation rules.

Read the original specification for food dynamics, energy costs, local sensing, sequential move semantics and the individual and competitive evaluation goals.

DECISION LOOP

A compact distributed brain

Every organism combines local perception with an inherited policy and a tiny shared signal. The controller adapts its priorities to food density, energy and life stage, producing coordinated population behaviour from strictly local decisions.

food nearbyenergyage8-bit signal

→

LOCAL ESTIMATORenvironment bucket

scarcestablerich

→

THRESHOLD POLICY1 eat2 reproduce3 explore4 wait

DNA + stateoffspringupdated signal

02WHY IT IS HARD

Partial observability

Every brain sees only its immediate neighborhood, while the parameters that define food availability remain hidden.

Ecological balance

Movement finds resources but costs energy; reproduction grows the population but can trigger starvation through overcrowding.

Competition

Other species consume food, block movement and distort the observations used to estimate the environment.

03OUR APPROACH

One adaptive brain,
not a collection of scripts.

Our final Java controller, FinalBrain, uses one hierarchical decision framework whose thresholds change with estimated food availability, energy, age and competitive pressure.

01

Estimate

Infer local food-spawn probability from repeated observations and classify it into robust probability buckets.

02

Share

Encode species identity and environmental beliefs into external state, then aggregate trusted neighbors' estimates.

03

Adapt

Adjust movement and reproduction thresholds across lifecycle phases, balancing early expansion with later conservation.

04

Inherit

Pass accumulated trials, food hits and age to offspring through a compact 32-bit DNA representation.

04JAVA DECISION PIPELINE

Perceive. Learn.
Choose. Survive.

Every organism independently executes the same modular controller on each simulation step.

PerceptionRead food, neighbor states and current energy.

EstimationUpdate local trials, hits, probability bucket and confidence.

ConsensusCombine trustworthy observations shared by nearby allies.

DecisionPrioritize food, reproduction, emergency search, exploration or waiting.

05ONLINE ESTIMATION

Learning food availability from censored observations.

The brain compares empty neighboring cells across consecutive turns. A valid empty-to-empty observation increments trials; if food appears, it also increments hits. The local estimate is therefore p̂ = hits / trials, discretized into behavioral buckets rather than trusted as a precise value.

Occupancy changes are excluded because another organism may have moved into the cell. This improves local validity, but competition introduces censoring bias: cells where food appears are also more likely to be consumed or occupied before they can become a positive observation, causing systematic underestimation.

LOCAL EVIDENCEempty_t → empty / food_t+1

trial observable transitionhit food appearedignore cell became occupied

The estimator is useful, but its assumptions change once competitors modify the environment being measured.

068-BIT COMMUNICATION

A tiny message carrying identity, belief and confidence.

Neighboring organisms expose only one 8-bit external state. We divide it into a species tag, the estimated probability bucket and a confidence field. The tag filters unrelated species; trusted messages are accumulated into a weighted local consensus.

Each bucket receives support proportional to confidence, including the organism's own estimate. The best-supported bucket controls behavioral thresholds, allowing local observations to become a distributed approximation of the global environment.

EXTERNAL STATE

3 bits species tag3 bits p bucket2 bits confidence

[ TAG ][ BUCKET ][ CONFIDENCE ]

07LIFECYCLE CONTROL

The same brain changes priorities over time.

FinalBrain maps environment bucket and age to thresholds for waiting, reproduction and movement. Rich environments encourage early spreading and growth; sparse environments begin cautiously and become more exploratory as survival pressure increases.

Energy constraints can override the normal policy: food takes priority, reproduction requires a safe reserve, and critically low energy activates emergency search. Modesty rules also suppress reproduction in crowded neighborhoods to avoid local resource collapse.

EARLYClaim space

Explore or farm food according to the inferred environment.

MATUREBalance

Reproduce only with sufficient energy and local capacity.

LATEConserve

Reduce risky movement and preserve a stable population.

08KNOWLEDGE INHERITANCE

A warm start for the next generation.

New organisms should not need to relearn an environment from zero. During reproduction, the parent packs its accumulated evidence into a 32-bit integer passed to the child.

32-BIT DNA

12 bits trials12 bits hits8 bits age

(trials << 20) | (hits << 8) | age

09MY CONTRIBUTION

Designing the adaptive core.

I developed BasicPlayerMD1 and BasicPlayerMD2, introducing the hierarchical rule-based architecture that became the foundation of FinalBrain.

I proposed and implemented local estimation of the hidden food-spawn probability p, its discretization into probability buckets, and the idea of building a global approximation through distributed information sharing and consensus.

I also designed and implemented the DNA-based inheritance mechanism, led much of the core implementation and contributed substantially to the strategic direction, report and presentation.

10RESULTS

Excellent alone.
More fragile in competition.

FinalBrain validated the value of energy-aware exploration and controlled reproduction in individual trials. Competitive play exposed slower adaptation and recovery when opponents distorted observations or occupied territory aggressively.

4,686.84Average final population in Big, ranked 1st with no extinctions.

1,901.99Average population in Finding Gold, ranked 1st.

174.28Average population in Default, ranked 2nd and 1st in final energy.

LessonOne adaptive policy reacted more slowly than specialized competitive strategies.

11WHAT WE LEARNED

Learning can be biased

Competitors consume food and alter the observations used to infer the natural environment.

Stability has a cost

Conservative reproduction avoids collapse but can make recovery and territorial expansion too slow.

Scale changes engineering

Thousands of organisms magnify allocations and repeated computations, making caching and simpler hot paths essential.

APPROXIMATE ≠ ARBITRARY

06 / ALGORITHMS & EDUCATION

Can We Cheat NP-Hardness?

A theoretical study of parameterization and approximation, paired with an interactive laboratory for TSP algorithms, empirical experiments and real road networks.

Complexity TheoryTSPVisualizationFull-stack

06 / THEORY + SOFTWARE · SPRING 2026

Can We Cheat
NP-Hardness?

A study of algorithms, illusions and reality: two theoretical ways to make hard problems manageable, and a full interactive laboratory that tests where those ideas survive contact with practice.

Complexity theoryPython + FlaskNumPyCanvasLeaflet + OSRM

View code on GitHub ↗

WORST CASENP-hard restrict the parameter exploit structure k△

CENTRAL CLAIM Hardness is a worst-case characterization, not always a practical verdict. The relevant question is what structure we are allowed to use.

01THE QUESTION

Why do we routinely solve problems that theory calls intractable?

Routing, scheduling and combinatorial optimization are formally NP-hard, yet useful solutions are produced every day. This project investigates that tension rather than treating NP-hardness as the end of the discussion.

The report develops two complementary mechanisms. A hard problem may become tractable when a parameter is small, or approximable when the input has additional structure. These mechanisms expose an important asymmetry: theory can be too optimistic about algorithms that are polynomial but impractical, and too pessimistic about approximation bounds that are rarely approached on real instances.

02TWO LENSES

PARAMETERIZED COMPLEXITY

Longest Path through Color-Coding

Randomly assign k colors to the graph, then search only for a path whose vertices have distinct colors. A subset dynamic programme replaces the global “all vertices must be distinct” condition with local color-set transitions.

one iterationO(m · k · 2^k)

STRUCTURED APPROXIMATION

Metric TSP through triangle inequality

General TSP resists constant-factor approximation, but metric structure makes shortcutting safe. MST doubling yields a 2-approximation; matching odd-degree vertices improves the guarantee to Christofides’ 3/2.

guaranteeℓ(C) ≤ 3/2 · OPT

03COLOR-CODING

Turn vertex identity into a palette.

A fixed path becomes colorful with probability k!/k^k ≥ e^−k. Repeating the experiment λe^k times pushes failure below e^−λ, without introducing false positives.

one-sided Monte CarloDP[v, S]confidence via repetition

DP STATEDP[v, {red, blue, gold, green, purple, coral}] = trueeach color appears exactly once

04CHRISTOFIDES

A 3/2 guarantee built from two bounds.

For a complete metric graph, Christofides first computes a minimum spanning tree T. The vertices of odd degree in T are paired by a minimum-weight perfect matching M. Their union is connected and Eulerian; shortcutting its Eulerian circuit produces a Hamiltonian cycle without increasing cost.

ℓ(T) ≤ OPT+ℓ(M) ≤ ½ OPT⇒ℓ(C) ≤ 3/2 OPT

01Metric graphcomplete + triangle inequality

→

02MST + odd setO = {v : degT(v) odd}

→

03Perfect matchingH = T ∪ M is Eulerian

→

04Eulerian circuitHierholzer on H

→

05Shortcut repeatstriangle inequality ⇒ no increase

05THE LABORATORY

Three scales of understanding.

I built a web laboratory where the same theory can be inspected as a single execution, measured statistically and challenged on real road data.

MODE 01 · MICROSCOPIC

Single Instance Study

Generate 5–30 Euclidean checkpoints and play four algorithms frame by frame. Every MST edge, matching, Eulerian traversal and individual shortcut receives synchronized mathematical commentary.

Prev / NextAuto-playPNG snapshot

MODE 02 · STATISTICAL

Empirical Simulator

Run up to 350 independent trials, compute OPT exactly for n ≤ 20 and watch running means for wall-clock runtime and approximation ratio converge live.

Live chartsα = ℓ / OPTCSV export

MODE 03 · GEOGRAPHIC

Road-Network Study

Place checkpoints on a real map, fetch directional OSRM distances and expose asymmetry, triangle violations and the “symmetrisation tax” between the model and the driven route.

LeafletOSRMGPX + Maps

06ENGINEERING

Minimal moving parts, visible algorithmic state.

The backend is the single source of truth: it computes complete step descriptors containing nodes, edges, commentary and statistics. The browser remains presentational and can render the same state onto either a Canvas or a geographic map.

Single instanceSimulatorRoad map

JSON

FLASK BACKENDalgorithms + step descriptors/generate · /run · /sim_round · /map_solve

HTTP

NumPyHeld–Karp kernelOSRMdistance + route

~200×Held–Karp speed-up at n = 20 after replacing interpreted loops with a vectorized broadcast-and-reduce kernel.

800 × 600Logical Canvas coordinate system, scaled for responsive and retina displays.

6 endpointsA compact API connecting generation, execution, simulation and real-map analysis.

07ALGORITHMIC KERNELS

01Nearest Neighbor

Fast greedy baseline with no constant guarantee, useful for exposing the gap between simple heuristics and theory.

02MST 2-Approximation

Prim, edge doubling, Hierholzer and shortcutting, animated one construction at a time.

03Christofides–Serdyukov

MST plus odd-vertex matching. The implementation uses an explicit, auditable greedy matching and documents the resulting loss of the strict 3/2 guarantee.

04Nearest Insertion

An O(n²) incremental tour builder that selects the closest outside vertex and cheapest insertion point.

05Held–Karp

Exact O(n²2ⁿ) dynamic programming used as the OPT reference for every ratio up to n = 20.

06Road diagnostics

O(n³) triangle scans, asymmetric-pair detection, symmetrization and directional re-pricing.

08THEORY MEETS DATA

The 3/2 line is a boundary, not a prediction.

On random Euclidean instances, the empirical simulator consistently places Christofides far below its worst-case bound. The important engineering choice was to compare every heuristic against an exact Held–Karp optimum rather than against another approximation.

1.00 OPT1.251.50 guarantee

Christofides1.05–1.10

MST 2-Approx.1.10–1.18

Nearest Insertion≈ 1.10–1.18

Nearest Neighbor1.15–1.25

09REAL ROADS

A → BB → A

The textbook assumptions break on a city map.

One-way streets and constrained turns make driving distances asymmetric. Some triples even violate triangle inequality. The application detects both phenomena, symmetrizes the matrix to run Christofides, then re-prices the resulting ordering in the actual travel direction.

THE MOST INFORMATIVE OUTPUTsymmetrisation taxdirectional cost − modelled cost

10CHOICES & LIMITS

Every simplification is made explicit.

The laboratory is designed for understanding and experimentation, not disguised as a production route optimizer. Its boundaries are part of the result.

01 / MATCHING

Greedy, not blossom

The implementation chooses an auditable greedy matching. It performs strongly empirically, but the strict Christofides 3/2 proof requires a true minimum-weight perfect matching.

FORMAL EFFECT · bound may rise to 2·OPT

02 / EXACT REFERENCE

OPT stops at n = 20

Held–Karp gives an exact comparison point, but its O(n²2ⁿ) state space imposes a hard practical ceiling even after vectorization.

NEXT STEP · branch-and-bound or 1-tree relaxation

03 / ROAD MODEL

The map solves a projection

Christofides optimizes a symmetric projection of the directional OSRM matrix. Re-pricing measures the tax, but does not turn the result into an ATSP guarantee.

HONEST OUTPUT · modelled cost vs driven cost

04 / DATA

Road estimates are a baseline

Public OSRM reflects static OpenStreetMap attributes rather than live traffic, closures or time-dependent travel conditions.

PRODUCTION PATH · self-hosted or live routing service

11WHAT I BUILT

A single argument expressed as proof, software and experiment.

I developed the theoretical narrative and proofs, designed the visual language of the laboratory, implemented its algorithmic kernels and full-stack architecture, built the statistical and geographic modes, and documented both the mathematical ideas and the engineering trade-offs.

MIHNEA / WELCOMEIdeas, systems
& meaningful impact.

07 / PERSONAL SOFTWARE

This Website

A portfolio designed as an explorable personal computer: folders, project windows, embedded documents and an AI guide turn a static profile into an interactive system.

Product DesignFrontendInformation ArchitectureResponsive UI

07 / PERSONAL SOFTWARE · 2026

A portfolio
you can explore.

Instead of arranging my work as a conventional sequence of pages, I built a small personal operating system: a place where visitors can browse ideas, open projects, read documents and ask for guidance.

HTMLCSSJavaScriptResponsive designAI interface

MSDPortfolio OS

Research Engineering Journal

Mihnea / Welcome

START_HERE.MIHNEA Ideas, systems
& meaningful impact.

Computer science, education, markets and public life.

SolvyBETA

Where should we begin?

research · projects · teaching

↖

01THE IDEA

Make personality part of the interface.

A traditional portfolio tells visitors where to scroll. This one lets them choose a path. The desktop metaphor makes a large and varied body of work feel familiar, while the visual language connects computer science with a more personal, human presentation.

Not a résumé rendered online, but a small world with its own logic.

02EXPERIENCE ARCHITECTURE

01Enter

A calm desktop introduces the person before the credentials.

→

02Choose

Folders divide the work by purpose rather than chronology.

→

03Explore

Dedicated windows hold detailed case studies and PDFs.

→

04Ask

Solvy offers a second, conversational path through the same information.

03WHAT IT DOES

WINDOW SYSTEM

Projects stay in context.

Case studies open like applications, preserving the desktop and making deep exploration feel lightweight.

PDF

LOCAL DOCUMENTS

PDFs belong inside the site.

Reports, statements and the CV use a reusable internal reader instead of repeatedly sending visitors elsewhere.

CONTENT LIBRARY

Teaching can grow naturally.

A file-explorer model supports nested folders and new public resources without redesigning the page.

RESPONSIVE SYSTEM

The metaphor survives mobile.

Windows become full-screen workspaces, navigation simplifies and typography scales without losing the visual identity.

04AI-NATIVE NAVIGATION

SOLVY / BETA

Some visitors know what they want.
Others only know what they are curious about.

Solvy is designed as a guide rather than decoration. It maps natural questions to the relevant part of the portfolio, creating an alternative to menus and folders. The current deterministic version establishes the interaction model; a richer retrieval-based assistant can later answer from the site's complete content.

VISITORWhat kind of research does Mihnea do?

SOLVYLet me open the work on private learning and algorithms.

05DESIGN PRINCIPLES

01

Clear before clever

The desktop should make the portfolio easier to navigate, never turn it into a puzzle.

02

Depth on demand

Short project summaries lead to detailed technical stories only when the visitor asks for them.

03

Honest incompleteness

Ongoing research, future work and unfinished sections are labelled rather than overstated.

TEACHING/MATERIALS

Resources built
to make ideas click.

Course material, exercise notes and resources from my teaching at ETH Zürich.

COMMUNITY/CIVIC LIFE

Pay attention.
Understand deeply.
Act with care.

I believe understanding the world around us is part of our responsibility toward it. Staying informed, questioning easy answers and helping where we can are small but meaningful ways of contributing to a healthier society.

This is not about having an opinion on everything. It is about remaining curious, resisting indifference and trying to leave the systems and communities around us a little better than we found them.

01 / CAUSES

Things I believe are worth caring about

01

Education & Opportunity

High-quality education should not depend on where someone is born or the resources of their family.

02

Children's Rights

Every child deserves safety, dignity, healthcare and the freedom to develop their potential.

03

Responsible Technology

Algorithms should serve people fairly, transparently and with attention to their social consequences.

04

Civic Participation

Healthy institutions depend on informed people who participate constructively and hold power accountable.

05

Economic Dignity

Markets are powerful tools, but prosperity should create genuine opportunity and protect human dignity.

06

Future Generations

Our decisions should account for the people who will inherit their environmental and institutional effects.

02 / INFORMATION

How I try to stay informed

Being informed is less about consuming more content and more about developing better habits of attention.

01
Read beyond headlines and look for the original evidence.
02
Compare serious sources with different perspectives.
03
Distinguish reporting, analysis and opinion.
04
Choose long-form work over a permanently reactive feed.
05
Remain willing to update a belief when the facts change.

03 / READING THE WORLD

A small, evolving bookshelf

Algorithmic ethics

The Ethical Algorithm

Michael Kearns & Aaron Roth

Fairness, privacy and socially aware algorithm design.

Search & society

Algorithms of Oppression

Safiya Umoja Noble

How search engines can reproduce and amplify social bias.

Markets & morality

What Money Can't Buy

Michael J. Sandel

Where markets belong, and where they may crowd out important values.

Justice

Michael J. Sandel

An accessible journey through competing ideas of a just society.

Institutions

Why Nations Fail

Daron Acemoglu & James A. Robinson

How political and economic institutions shape prosperity.

AI & society

The Alignment Problem

Brian Christian

The human values hidden inside machine-learning systems.

Education

The Smartest Kids in the World

Amanda Ripley

What different education systems can teach us about learning.

Evidence

Factfulness

Hans Rosling

Better instincts for interpreting global trends and statistics.

04 / ORGANISATIONS

Work worth discovering

Organizations whose work I follow and encourage others to explore. Inclusion here is an invitation to learn more, not an affiliation or endorsement of every position.

GLOBAL

Large international organizations working across borders.

UN

Children

UNICEF

Protecting children's rights, health, education and development worldwide.

↗ SC

Children

Save the Children

Supporting children's education, protection and wellbeing around the world.

↗ KA

Open education

Khan Academy

Making high-quality learning resources freely available to anyone.

↗ ICRC

Humanitarian action

International Committee of the Red Cross

Protecting life, dignity and access to humanitarian assistance worldwide.

↗

SWITZERLAND

Work supporting young people and responsible technology in the community where I live.

PJ

Children & youth

Pro Juventute

Supporting children, young people and families across Switzerland.

↗ AW

Responsible technology

AlgorithmWatch CH

Examining algorithmic systems and promoting transparent, accountable technology.

↗

ROMANIA

Civil-society initiatives addressing education, children and public infrastructure.

SC

Children

Salvați Copiii

Education, protection and healthcare for vulnerable children.

↗ TF

Education

Teach for Romania

Reducing educational inequity through teachers and leadership.

↗

UNITED STATES

Organizations using education and technology to widen opportunity.

DC

Public education

DonorsChoose

Helping public-school teachers fund resources for their classrooms.

↗ CO

Computer science education

Code.org

Expanding access to computer science education in schools.

↗

THINKING OUT LOUD · ZÜRICH

Ideas are clearer
once they are written.

A quiet place for unfinished questions, lessons from practice and longer reflections across computer science, education, markets and public life.

WHAT MAY APPEAR HERE

I.

Ideas

Technical questions and concepts I want to understand more deeply.

II.

Field Notes

Lessons gathered while researching, teaching and building things.

III.

Reflections

Longer thoughts on education, technology, markets and society.

Mihnea/Engineering/Parallel Football

GitHub ↗

Mihnea/Engineering/Mosquito

GitHub ↗

Mihnea/Engineering/Organisms

GitHub ↗

Mihnea/Engineering/Soldier, Soldier, Soldier, Spy

GitHub ↗

Mihnea/Engineering/Can We Cheat NP-Hardness?

GitHub ↗

Mihnea/Engineering/This Website

LIVE SYSTEM

Mihnea/Engineering/Local

ANDROID · V0.1.0

Mihnea/Engineering/TSP Interactive Laboratory

GitHub ↗

INTERACTIVE LAB / DEPLOYMENT

The window is ready.
The backend needs a home.

The visualizer runs on Flask and NumPy, so it must be deployed as a small web service before it can live inside this window. Once deployed, this page will load the complete application here without sending visitors outside the portfolio.

RecommendedRender or RailwayFlask service → embedded securely in this desktop window

View the application source ↗

Mihnea/CV.pdf

Download ↓

I make sense ofcomplex things.

Ideas, systems& meaningful impact.

Let's talk!

Computer Science

Education

Public Life

Markets

Leadership

Science & Discovery

Bucharest

Zürich

Philadelphia

Amsterdam

Coming soon

Can mixing create privacy?

Questions I have been following

Private PAC Learning

Samples vs. Computation

Geometry & Optimization

Programming & Problem Solvingin the Age of Generative AI

A new theoretical question,still taking shape.

Local

A social graph that beginsin the real world.

Rich conversations

Ephemeral by choice

Useful, not bare

Sends that survive navigation

Privacy expressed through product decisions.

Two scans. Two ephemeral keys.One shared relationship.

A new message keyfor every message.

From tap to trusted local state.

Large files use envelope encryption.

Native, local-first architecture.

Atomic ratchet state

Granular recomposition

Fast chat opening

Durable asynchronous work

Explicit data evolution

Release-oriented build

Controlled session ownership

Typed control events

Operational safeguards

Privacy continues after decryption.

Privacy also means deleting correctly.

Conversation content

Routing metadata

Built seriously.Described honestly.

Programming &Problem Solving

Mosquito

Spy

Football

Organisms

Corral a stochastic population with a handful of lights.

The complete simulation rules.

One lattice, many formations

Coupled decisions

Noisy evaluation

Finite search time

Eight iterations, each removing a mistaken assumption.

Discretize first.Recover continuity later.

Spend precision only where it matters.

A graph coloring becomes a control policy.

When search quality became a throughput problem.

Owning the core solution end to end.

Strong ideas, uneven robustness.

Representation changes the problem

Measure the evaluator

Test the hard regime

A race for 1,020 footballs.

Read the complete original statement.

Coordinate a moving field

The rules are local.The consequences are global.

Contention

Coordination

Different regimes

Start simple.Measure. Adapt.

BasicPlayerMD

GreedyPlayer3

TwiceImprovedPlayer

HybridPlayer3

I make sense of
complex things.

Ideas, systems
& meaningful impact.

Programming & Problem Solving
in the Age of Generative AI

A new theoretical question,
still taking shape.

A social graph that begins
in the real world.

Two scans. Two ephemeral keys.
One shared relationship.

A new message key
for every message.

Built seriously.
Described honestly.

Programming &
Problem Solving

Discretize first.
Recover continuity later.

The rules are local.
The consequences are global.

Start simple.
Measure. Adapt.

Evidence is not just a value.
It has a history.

Correctness first.
Progress before timeout.

One adaptive brain,
not a collection of scripts.

Perceive. Learn.
Choose. Survive.

Excellent alone.
More fragile in competition.