Selected work
9 real, shipped projects, each written up as a full case study covering the problem, the engineering decisions, the results, and what did not work. Ordered from most recent.
A 14M-row randomized ad experiment that clears every statistical bar and still loses money, until an X-Learner and a profit-aware bandit find who to actually target.
A 9.5M-row UK supermarket price pipeline rebuilt from a broken prototype into a forecasting model, an explainability layer, and a live dashboard.
A 109M-event e-commerce behavioral pipeline built to run on a laptop: DuckDB and Polars instead of a warehouse, RFM segmentation, a propensity model, and a dashboard, then rebuilt for reproducibility after two silent non-determinism bugs surfaced under audit.
12 hand-written CUDA kernels for LLM, MoE, and RAG workloads, wrapped in Rust and exposed to PyTorch as zero-copy ops, three of them proven inside a real Llama-3-8B block.
A from-scratch Rust and CUDA reimplementation of vLLM-style paged KV-cache memory: page GPU memory on demand instead of pre-allocating the worst case, benchmarked against HuggingFace's own cache.
An institutional-grade SEC 13F research pipeline audited by actually running it on 120M real filings: 12 real bugs found and fixed, including a signal engine that could not run at all, then re-run for real, honestly-reported results.
A LangGraph agent that plans, retrieves, and reasons over SEC filings, then checks its own draft for hallucination before answering, engineered to survive Groq free-tier rate limits without falling over.
A genetic algorithm searches 3.1x10^10 layer-wise quantization configs for a real LLM, then checks its own proxy against real GPU hardware instead of assuming it works.
A 358,497-row PySpark batch pipeline and three scikit-learn models behind a Streamlit dashboard, audited by actually running the shipped models: the dashboard's headline accuracy numbers trace to a model that was never deployed.
← Back to home