Projects.

Neural-operator surrogates for Monte Carlo particle transport, trained on cheap, noisy MC labels across 1M+ configurations. Keeping supervision in linear space behind a positivity-preserving head sidesteps the Jensen bias of log-transformed noisy labels, so the model predicts fields spanning 9–11 orders of magnitude, ends up more accurate than its own training labels, and runs inference up to ~10⁶× faster than converged Monte Carlo. Work with Prof. Anima Anandkumar's lab at Caltech.

A multi-agent system that simulates real user populations with swarms of LLM agents to replace traditional A/B testing, running thousands of agents concurrently per simulation. I own the stack (Next.js, AWS, Temporal) and the model-selection and cost trade-offs against real usage. In production for 10+ paying design and brand-naming studios at $5K MRR, on a two-person team.

Can supervised fine-tuning teach an LLM which headline gets clicked? LoRA SFT of Qwen3-14B on the Upworthy Research Archive made ranking worse, twice, while eval loss and token accuracy kept improving. The lesson: checkpoint on the capability you care about, not its proxy.

An AI assistant on Snapchat Spectacles that streams the wearer's view to Gemini Live and guides circuit and electronics repair with voice and AR overlays.