ls ~/work
Nearly everything I do circles back to decision-making under uncertainty: quantitative trading and markets, Bayesian modeling, and getting language models to know what they don't know. Consider this a running index of all of it – the research, the work, and the side projects alike.
BAIR research on language models that stay fluent when they shouldn't. Sample summaries, build a probability-weighted NLI entailment graph to localize the unstable proposition, and ask one targeted question when calibrated uncertainty crosses threshold. Pending submission to ACL Rolling Review.
Graduate Bayesian statistics project benchmarking KSC Gibbs, NUTS, and flowMC samplers: mixing, efficiency, and posterior calibration across classical MCMC and normalizing flows.
In Rich Ivry's Cognition and Action Lab: recurrent networks learning to reach for targets under rotated and mirror-reversed visual feedback. Built custom Gym environments with configurable perturbation dynamics and probed the GRUs' latent dynamics against human adaptation curves.
At MyMagic AI: a 7B draft model proposing tokens for a 70B verifier inside a custom vLLM fork. 2.6× speedup, 35 to 90 tok/s at a 78% draft acceptance rate, output equivalent to greedy decoding.
Head TA for Business and Finance for Engineers; helped shape it from the EE 194 pilot into a permanent course. Wrote most problem sets, including the derivatives, options, and personal finance ones, built the PrairieLearn and PrairieTest infrastructure and grade report generation, and wrote a reusable LaTeX package so the course outlives me, plus speakers from Sequoia and the finance world.
Sole contract engineer on a document intelligence platform for title insurance: DocumentAI OCR, NER-guided correction, human-in-the-loop validation, and a RAG layer drafting commitments at over 90% field-level accuracy. Cut examiner effort 96%; piloted by 20+ examiners across three insurers. Nothing too glamorous, but it's honest work.
Founding engineer on a sports prediction startup. Built a TensorFlow ensemble of XGBoost and neural nets with 23 engineered features for NFL over/unders: 64% accuracy and 4.2% ROI across 400+ games, with Monte Carlo backtesting over 10k+ scenarios and Kelly sizing to separate real edge from variance. Accepted to Y Combinator's X25 batch, but the founding team went separate ways.
A kid who'd just learned what an LSTM was, with a good mentor. It backtested around 5% annualized excess returns and got used by half a dozen analysts making real allocation calls. The unglamorous part I'm proudest of: a pipeline parsing 1,000+ semi-structured statements into 50+ IFRS-aligned features.
Top 1% globally in IMC's trading competition. Found a basket-sum invariant and cross-product lead-lag, priced options with Black-Scholes, and gamed the manual rounds by mining competitor Discord chatter into a Gaussian mixture over the field's likely picks.
Plain-English trading for Hyperliquid perps, built for the Trading Track at BaB Hacks 2026. You say or type an order and it parses your intent, checks the live market, places the trade, and logs your thesis, then calls you out when you start contradicting yourself. Best of all is Jim, the voice agent who runs the whole thing by phone and rings you when a position drifts toward liquidation.
Mostly built on vibes.
Implemented n-gram language models with smoothing alongside a feed-forward neural language model, comparing count-based and neural next-token prediction.
Implemented a Neural Radiance Field from scratch: an MLP learning density and color for 360° view synthesis.
Built diffusion sampling, CFG editing, inpainting, and illusions, then trained a UNet via flow matching on MNIST.
Built a panorama pipeline: feature detection, descriptor matching, RANSAC homographies, and multi-band blending.
Treated images as functions: filtering, sharpening, hybrid images, and frequency-based blending.
Digitally reconstructed and color-corrected glass plate photographs from the Prokudin-Gorskii collection.
Exploring how focal length, perspective, and camera position shape images.
Built neural nets and attention in PyTorch: digit classification, language identification, GPT-style generation.
Built exact and particle-filter inference tracking hidden ghosts with Bayes nets and HMM time-elapse models.
Implemented Q-learning and policy gradient methods to train Pac-Man agents.
Implemented Minimax, Alpha-Beta pruning, and Expectimax agents for Pac-Man with stochastic ghost models.
Built DFS, BFS, UCS and A* search agents with custom heuristics to guide Pac-Man through mazes efficiently.
Built a fully functional RISC-V CPU supporting all 7 instruction types; verified with custom test benches.
Feedforward neural net with matrix multiply, ReLU, argmax in RISC-V.
Implemented 2D convolutions, then accelerated with AVX2 SIMD, OpenMP threading, and an Open MPI coordinator.
Built a playable Snake engine in C. Managed game state, rendering, and user input on a terminal interface.
Built a Transformer language model end to end - self-attention, the training loop, and sampling-based decoding.
Built a RAG system answering factoid questions about Berkeley EECS: an 18,706-chunk crawl of eecs.berkeley.edu, dense FAISS and sparse BM25 retrieval fused with reciprocal rank fusion, and few-shot Llama 3.1 8B generation, reaching 61.76% F1 on the held-out test set.