MRes @ Imperial (ML & Big Data) · BSc Theoretical Physics @ UCL · Currently making GANs hallucinate particle beams.

I sit at the intersection of physics, ML, and software engineering - which mostly means I spend a lot of time staring at loss curves and pretending I understand why they diverge. My background is in theoretical physics, but I got tired of maths that doesn't ship.

• 🧲 Right now: Building a WGAN-GP to replace Monte Carlo particle simulations for the COMET experiment. Targeting a million-fold speedup. It's going fine.
• 🤖 Day job: Evaluating LLM failure modes at Google; turns out even the best models have bad days.
• ♠️ Previously: Semi-professional poker player. Managed a five-figure bankroll for two years. Poker and transformers are the same problem: large-scale probability estimation over sequences, where edge comes from better calibration than the competition.
• 🏐 Unrelated: National gold medallist in volleyball. Coached two teams to league promotion this year.

The fun stuff

The necessary stuff

The impressive-sounding stuff

WGAN-GP that learns the phase space of 105 MeV pion→muon decays for the COMET experiment, replacing Geant4 Monte Carlo tracking. Running on Imperial's HPC cluster via HTCondor. Projected speedup: ~1,000,000×. Physics-informed constraints baked in to keep conservation laws intact.

Custom CNN (405k params, ~1.5MB) built for UAV imagery — ~94% recall. Then compressed it: L2-norm structured pruning + INT4 quantisation, deployed to FPGA via ONNX and HLS4ML. The full pipeline from training to on-chip inference.

Benchmarked long-context compression strategies (sliding window, summarisation, hybrid) across 36 synthetic multi-turn conversations. Framed as an optimisation problem: push compression until performance breaks, then understand why. Ruled out turn-selection by similarity — unrealistic for live compression.

Full-stack RAG system for high-energy physics papers. Streaming FastAPI backend, vector search via ChromaDB, local LLM inference via Ollama. Fully containerised. No cloud, no data leaving the machine.

Q-learning agent trained on 250,000 real-money hand histories. Achieved +8bb/100 in simulation. Automated hand history ingestion from PokerTracker. Borderline irresponsible amounts of variance analysis.

Email: jonathancassens@gmail.com · LinkedIn: jonathan-cassens

English (native) · German (native)