Ilyes Jamoussi

Selected projects

1. Peaked Circuits

   code↗ (opens in new tab)

   A quantum-advantage paper, reproduced number for number.

   • Reproduction of Aaronson & Zhang (arXiv:2404.14493), Section 3: 1D brick-wall circuits — 40 Haar-random layers followed by 10 variational SU(4) layers trained to concentrate the output distribution on a single basis state.
   • Peak weight maximized with batched Adam restarts, step-decay learning rate and early stopping; every run is fully seeded and the saved artifact rebuilds the exact circuit.
   • Reaches peak weight 0.2024 at 12 qubits against the paper's reported ~0.2, boosting the peaked string three orders of magnitude above the 2⁻¹² baseline.

   δ = 0.2024 (paper: ~0.2) · 12 qubits · 50 layers · CI-testedPennyLane · NumPy · SciPy

   

2. MiniGPT

   code↗ (opens in new tab)live demo↗ (opens in new tab)

   A GPT built from scratch, down to the tokenizer.

   • Decoder-only GPT in pure PyTorch — causal multi-head attention, learned positional embeddings, weight tying — with no nn.Transformer and no pretrained weights.
   • Byte-level BPE tokenizer (8,192-token vocabulary) trained on the train split only; 13.9M parameters trained on TinyStories to 5.32 validation perplexity.
   • Token-streaming inference over FastAPI (server-sent events), Dockerized, with CI running pytest, ruff, and mypy.

   “Once upon a time, there was a little girl named Lily. She loved to play outside in the park. One day, she saw a big, juicy apple on a tree. She picked it up and ate it all up.”

   13.9M parameters · 5.32 val perplexityPyTorch · FastAPI · Docker · Streamlit

3. MailGuard AI

   code↗ (opens in new tab)live demo↗ (opens in new tab)

   Transformer spam detection, every layer hand-built.

   • Encoder-only Transformer in PyTorch — padding-masked multi-head attention, sinusoidal positional encoding, masked mean pooling — implemented entirely from scratch.
   • Custom 30K-token vocabulary built from the training split only; class-weighted loss on stratified splits.
   • 97.0% accuracy and 0.95 spam F1 on the SpamAssassin corpus.

   97.0% accuracy · 0.95 spam F1 · 10.9M parametersPyTorch · Streamlit

   Confusion matrix — 864 unseen test emails

   	pred. ham	pred. spam
   actual ham	600	16
   actual spam	10	238

4. Poly Arena

   code↗ (opens in new tab)live demo↗ (opens in new tab)

   Cross-platform multiplayer tactics game.

   • Cross-platform multiplayer tactics game — Angular web, Flutter mobile, and Electron desktop clients driven by one authoritative NestJS + Socket.IO server (MongoDB, Firebase Auth); 4 game modes, visual map editor, AI players.
   • 980+ unit tests (Jest & Jasmine), all passing in a GitHub Actions pipeline that tests, lints, and builds on every push.
   • Team project of 6 at Polytechnique — top contributor with 283 of 545 commits; led the Angular client and core server gateways.

   980+ unit tests · 3 platforms, 1 serverAngular · NestJS · Socket.IO · Flutter · Electron · TypeScript · MongoDB

5. AVR Robot

   code↗ (opens in new tab)

   Autonomous navigation in bare-metal C++.

   • Autonomous line-following robot in bare-metal C++ on an AVR ATmega324PA — self-localizes by matching its recorded move sequence against known route signatures and stops 10 cm from target posts.
   • Register-level PWM motor control, infrared sensing, I2C EEPROM persistence, and UART reporting — no Arduino, no HAL.

   C++ · AVR