A drug-gene interaction checker built on CPIC guidelines that refuses to guess. Its core rule: never let “no known interaction” and “cannot assess this gene” collapse into the same answer, because absence of data is not absence of risk. 452 tests.

A simulated exchange with a C++ matching engine on a zero-allocation hot path: ~61ns order adds, ~118ns limit matches, 8.5M matches/sec. Wrapped with pybind11 so RL agents can train against it through a Gymnasium env, with a live React dashboard over WebSocket.

Built for tax preparers doing year-over-year workpaper review: diffs two workbooks cell by cell with a materiality threshold, then reconciles the filed PDF return against the workbook that produced it, handling the sign flips that accounting actually uses.

Upload any document, get a typed REST API. Infers a schema from the PDF itself instead of making you configure templates or regexes up front, so it keeps working when the document format shifts.

Open pull requests under review on Starlette, OpenTelemetry Python, and Supervisor: an IPv6 host header parsing bug, a multi-byte decode crash, and missing real-time signals. Mostly an exercise in reading large unfamiliar codebases well enough to fix something narrow without breaking anything.

an MLIR-based optimizer for quantized transformer LLMs, aimed at edge deployment where you can't just reach for a bigger GPU. TensorRT and ONNX rewriting took 48% off latency on NVIDIA hardware.

a multithreaded engine handling 1,000+ datapoints/sec off live feeds, event-driven so strategy logic reacts to the market instead of polling a timer. VWAP execution and limit orders, benchmarked against the S&P and sector indices.

async LLM inference on serverless Lambda, where cold starts are the entire problem. caching plus streaming and rate-limiting layers cut response time 30%, at 98.4% interpretation accuracy.

forecasts S&P 500 movement over 30-day windows from macro and firm-level data, ~92% accurate. Snowflake and Airflow run the daily ETL, Tableau renders the output.

predictive models running over real-time market data pipelines, projecting ~$150 PnL/min at 95% accuracy.

pulls scattered patient data sources into one place and runs predictive analytics for early disease detection, with real-time dashboards on top.