Software engineering and machine learning research.

Owned design, testing, rollout, and documentation of shared configuration infrastructure for Amazon SCOT’s long-term supply-chain forecasting.

• 17 packages + services
• 30+ cross-functional users
• ~90% less manual comparison

Used by engineers, managers, economists, and scientists. Automated dependency tracing from research model outputs to downstream configurations; the reduction in manual comparison time is based on user surveys.

Shipped version-controlled JSON configuration delivery through Python, AWS Lambda, S3, and internal APIs, with infrastructure as code in TypeScript using AWS CDK.

Python, TypeScript, AWS Lambda, S3, CDK

Built and launched an open-source earnings analytics platform with searchable calendars, watchlists, and point and quantile forecasts from six LightGBM models.

• 120k+ earnings records
• 10k+ ticker identities
• 120+ monthly active users

Four-fold walk-forward validation with a five-day purge measured 31% lower weighted mean absolute error than the options-implied baseline on development data. Forecast pipelines include data-freshness checks and model-drift monitoring.

Nightly ingestion and batch inference run on DuckDB/Parquet and GitHub Actions. FastAPI serving uses signed artifacts, SHA-256 verification, atomic activation, and recovery to the previous verified model; Redis shares live quotes.

Regression tests cover corrupt downloads, mismatched bundles, missing features, and crossed prediction quantiles. Rejected model updates preserve the previously active model.

Next.js, TypeScript, Python, LightGBM, DuckDB, Parquet, PostgreSQL, Redis, FastAPI

Built an open-source financial research platform over SEC filings, combining PostgreSQL full-text search and pgvector approximate nearest-neighbor search with citation verification, evidence-based abstention, and point-in-time filtering.

• 3.04M embedded passages
• 3,204 SEC filings
• 75+ monthly active users

Multi-query expansion raised Recall@5 from 0.152 to 0.212 on a 33-query benchmark. Half-precision vectors reduced database size by 27%, from 15 to 11 GB, with no observed top-10 ranking change in benchmark comparisons.

Batched embeddings, approximate nearest-neighbor search, and bounded candidate pools cut 95th-percentile cross-company retrieval latency from approximately 59 seconds to 1.74 seconds in a retrieval benchmark.

Python, FastAPI, PostgreSQL, pgvector, Next.js

Led a four-person team building a video-to-inventory app with YOLO object tracking, cross-frame deduplication, Gemini valuation, and Supabase persistence.

• Top 5 of 172 teams
• 4 person team
• 2 category prizes

Integrated voice search and item snapshots into the inventory workflow. Placed Top 5 of 172 teams and won the Google Gemini and Financial Tech prizes at LA Hacks 2025.

Python, YOLOv11, Gemini, React, Flask, Supabase

Built a reproducible pipeline for studying deceptive communication in language models, linking 1,602 recipient-answer records to factual-check evidence and 990 decoder-layer activation matrices. The latest controlled comparison collected 288 answers and 144 separate factual checks across Qwen3-8B reasoning modes and Llama-3.1-8B prompt styles, with unchanged model weights. Preserved prompts, token positions, model revisions, and activation checksums for review. Labels remain provisional pending human validation; principal component analysis and detector training have not yet been performed on this dataset.

Built a NumPy research platform for Wasserstein and joint sliced-Wasserstein clustering across five exchange-traded funds, with chronological validation, robustness controls, frozen-model scoring, and resumable parallel experiments. An exploratory study of 600 five-session risk forecasts did not demonstrate an advantage over four simpler baselines.

Built an open-source Python prototype for tuning noisy simulations through parallel evaluation, adaptive replication, early stopping, and independent audits. In one 16-configuration synthetic experiment, it selected the same candidate as full-budget search using 72 rather than 432 search evaluations.

Developed a graph-convolutional and Informer model, evaluating four baselines across three horizons. Reduced 15-minute turning-ratio root mean square error by 24% versus STGCN on SUMO-simulated traffic data. Served as a conference Session Chair and was a Team Canada finalist for the International Science and Engineering Fair (Top 30).

Highest division of the USA Computing Olympiad, promoted from Gold with a 1000/1000 score.

Computational fluid dynamics and neural surrogate modeling for Formula 1 aerodynamics, achieving a 43% drag reduction.

Machine learning-assisted column generation for vehicle routing, producing a 15% faster C++ solver.