Both a technical showcase and an invitation to play

AI games & demos — to play with and to illustrate my skills

These mini-games are both meant to be fun and technical proofs of my AI/ML skills. Each demo implements a specific algorithm, model, or pipeline (Minimax, genetic algorithms, neural networks, computer vision, STT/translation, MediaPipe…) running locally in the browser — no server, no tracker, GDPR by design. → see all skills · the tags on each card map to the technologies actually used.

Filters and search

📖 New — technical case study of the VISION studio (context, decisions, failures, learnings) → Read the article

Chess Master

Strategy · Minimax

Full chess game with an interactive tutorial (10 lessons), adaptive AI across 7 levels and a Challenge mode that analyzes your habits to push you to improve.

  • Minimax
  • Alpha-Beta
  • PST
  • Adaptive
Play

GENESIS — The Spark of Life

Simulation · Boids + Genetic

A self-progressing evolution sandbox: six eras from quantum particles to a living ecosystem. Interactive entities, predation, symbiosis, genetic mutations.

  • Boids
  • Genetic
  • Procedural
  • Idle
Start the sim

RACEGEN — Living Track

Arcade · NN + Genetic

Race on a procedurally generated track. Drive yourself or watch an AI learn to drive through a neural network + genetic algorithm across generations.

  • Neural network
  • Genetic algorithm
  • Procedural
AI race

MAZE — Pathfinding Race

Demo · BFS · DFS · Dijkstra · A*

4 shortest-path algorithms launched in parallel on the same maze. Draw your walls, move the start and end points, watch the exploration waves spread and compare strategies.

  • Pathfinding
  • A*
  • Graph search
Start the race

VISION — Computer Vision Studio

Demo · Multi-model · TensorFlow.js

Analyze video, image or webcam with 10 AI modes: objects, faces + emotions, poses, scenes, anatomy, silhouette, motion, hands, license plates. 6 TF.js/WebGL models, before/after AI split screen, OCR, anonymization. 100% local — GDPR by design.

  • Computer Vision
  • 9 TF.js models
  • WebGL
Launch VISION

TRAIN — Live MLP fine-tuning

Demo · TensorFlow.js · MLP

Draw digits, label them, and watch a small MLP (784 → 128 → 64 → 10) learn in real time with TensorFlow.js. Loss and accuracy curves plotted at every epoch, live prediction via softmax. 100% local — GDPR by design.

  • TensorFlow.js
  • MLP
  • Backpropagation
  • Dataset UI
Launch TRAIN

PROMPT — Interactive prompt engineering

Demo · LLM · Local eval

10 challenges to sharpen the art of the prompt: JSON extraction, constrained summary, persona, few-shot, chain-of-thought, child safety guardrails, markdown table, bilingual. Heuristic local evaluation (keywords, format, structure, length). Score out of 100 + tips + reference prompt. Zero API, 100 % offline.

  • Prompt engineering
  • Few-shot
  • Chain-of-thought
  • Heuristic eval
Play PROMPT

VIZ — Interactive AI dataviz

Demo · KMeans · PCA · Regression

4 synthetic datasets (blobs, moons, spirals, noisy curve) and 3 algorithms in pure JS: KMeans (k-means++), 2D PCA (via exact covariance), polynomial regression (least squares). Sliders for hyperparameters, native canvas, step-by-step iterations. See convergence and overfitting live.

  • KMeans++
  • PCA
  • Least squares
  • Native canvas
Launch VIZ

RAG-PDF — Chat on a PDF, 100 % local

Demo · RAG · Transformers.js + pdf.js

Drop a PDF, the app extracts text via pdf.js, chunks it, computes an embedding per chunk with all-MiniLM-L6-v2 (Transformers.js), then retrieves the top-3 passages closest to your question by cosine similarity. Token highlighting. No server, everything in the browser.

  • RAG
  • Transformers.js
  • pdf.js
  • Cosine sim
Open RAG-PDF

SIGN — LSF Alphabet + Voice Translation

Demo · Accessibility · MediaPipe + Transformers.js

Real-time recognition of the 26 letters of the French Sign Language alphabet via MediaPipe Hands + rule-based classifier on 21 keypoints. Voice translation FR ↔ EN with Whisper-tiny (STT) + OPUS-MT + Web Speech API (TTS). 100% local — GDPR by design. Built-in spelling game.

  • MediaPipe Hands
  • Whisper-tiny
  • OPUS-MT
  • Accessibility
Launch SIGN

DEEPFAKE — Synthetic image detector

Demo · Computer Vision · ViT + Transformers.js

100 % local deepfake detection via Deep-Fake-Detector-v2 (ViT-base fine-tuned, ~86M params) on Transformers.js v3 with WebGPU backend. Three modes: image, live webcam and frame-by-frame video with score timeline. No server upload — GDPR by design.

  • Vision Transformer
  • Transformers.js
  • ONNX WebGPU
  • Image classification
Launch DEEPFAKE

SAM — Segment Anything, one click

Demo · Computer Vision · SlimSAM + Transformers.js

Object segmentation in one click via SlimSAM-77 (distilled × 8 Segment Anything Model, 77M params) on Transformers.js v3. One-time image encoding (~2s), then instant mask decoding (~200ms per click). Multi-point, exclusions, mask + cutout export. 100% local — GDPR by design.

  • SAM / SlimSAM
  • Transformers.js
  • Interactive
  • Segmentation
Launch SAM

2048 — Expectimax AI

Puzzle · Expectimax

Classic 4×4 2048, plus an Expectimax AI with adjustable depth that can play for you. Live visualization of per-direction evaluations.

  • Expectimax
  • Weighted heuristic
  • Tree search
Start the puzzle

Minesweeper — Logical AI

Puzzle · Constraints + Bayes

Classic Minesweeper with a solver that propagates the constraints of revealed numbers, then enumerates valid configurations to compute a per-cell mine probability (heatmap toggle).

  • Constraint propagation
  • Enumeration
  • Probabilities
Defuse

ADAPT — Adaptive Platformer

Platformer · Adaptive PCG

A 2D platformer where each level is generated based on your play style. A real-time profile (aerial, ground, aggressive, cautious, risky) weights the generator to deliver bespoke challenges.

  • PCG
  • Profiling
  • Weighting
  • Adaptive
Play

PULSE — Rhythm AI

Rhythm · Beat detection

Drop any audio file. A local AI isolates bass, detects onsets via energy + BPM autocorrelation, then drops notes in rhythm. Your music stays on your device — zero upload, zero persistence.

  • Onset detection
  • FFT
  • BPM
  • Privacy-first
Play

Why this library?

These demos are first and foremost a didactic showcase: each game illustrates a family of algorithms I work with daily (heuristics for decision-making, tree search, reinforcement learning, constraint propagation…). Everything is vanilla HTML/CSS/JS, no dependency, no server — matching the rest of the site (GDPR, no tracker, no remote fonts).