Work & projects

The agent starts with a 40M-parameter policy/value transformer trained by behaviour cloning on search-teacher data. I then train it with masked PPO and a KL anchor against a prioritised self-play league. Deck-specific LoRA specialists are trained on a frozen base and fused back into the model.

Decks evolve with MAP-Elites, using win rate as the fitness measure and a worst-case performance floor. At match time, the agent uses determinised tree search (PUCT) over a belief model of the opponent’s deck, within a two-CPU budget and without network access.

I evaluate changes with a tiered set of tests and power analysis. You can find the entry under “Rare Candy Lab 🍬” on the leaderboard.

I represented brain surface geometry with neural radiance fields, then optimised camera positions through backpropagation. The implementation is model-agnostic and built in nerfstudio.

I compared L1, L2, SSIM, NCC, and mutual-information losses for registration accuracy, convergence, and stability. The full thesis is “Enhancing Intraoperative Registration with Neural Radiance Fields: An Exploration of Loss Functions Effects.”

A ModernBERT classifier, fine-tuned on around 10,000 historical tickets, routes issues to components with 85% top-three accuracy.

The model training pipeline uses soft-verified-generation supervised fine-tuning following SERA, then on-policy reinforcement learning with SDPO (self-distillation policy optimisation). The aim is to replace frontier-model calls with task specialists at a cost below one dollar per fix, targeting roughly five times faster time-to-fix.

My teammates found the foundational research paper. I translated the approach into a prototype and wrote most of the code, connecting the iPhone’s magnetometer, accelerometer, and gyroscope to the mathematical methods and visualizations.

The demonstration reconstructed a walked path from inertial measurements and mapped the local magnetic field inside Station F. It brought sensor capture, magnetic-field scanning, navigation, and visualization together in one working prototype.

Adapting this to underwater drones was the proposed next step. We did not run underwater trials or validate the deck’s long-range accuracy estimates.

I worked on the software, backend, and data engineering. The team designed the system with the intention of making its code available for other groups setting up similar call centres.

We fine-tuned Wan2.2 to produce temporally stable videos, then trained a 4D Gaussian model to reconstruct and render the evolving scene. The project was called “API for the Generated Reality.”

We used NLP and vector search to retrieve sources for claims shared on social networks and elsewhere.

Built with LangChain, Next.js, Tailwind CSS, the ChatGPT API, and GCP. The project page includes the code, demo, and presentation.

The initial version used GPT-3, later GPT-3.5. We designed it as a short, personalised career test with a broad range of occupations and attention to data privacy.

We used a neural coverage fuzzing criterion in an existing framework and trained LeNet-5 on the German Traffic Sign Recognition Benchmark. The report examines the difficulties of estimating density and sparsity in high-dimensional spaces, including computational costs.

Full report: “Evaluating Test Data Quality by Measuring Latent Space Coverage, Density, and Sparsity with Deep Neural Networks.”

The paper covers physical simulations, fault localisation, transmission-system sensors, LiDAR, and remote sensing. It also proposes a three-layer approach to alternative power supplies in areas affected by shutoffs.