Projects

A full-stack fantasy sports platform built solo in seven weeks — league management, scoring, standings, and awards over live match data.

• TypeScript
• Next.js 15
• React 19
• PostgreSQL
• Drizzle
• Vitest
• GitHub Actions

• Built and shipped a solo full-stack platform in 7 weeks: 53,000 lines of TypeScript across a Next.js/React frontend, a framework-agnostic domain layer, 39 Postgres tables, 22 idempotent migrations, and 51 API routes.
• Designed an immutable-source, derived-everything architecture where standings, head-to-head results, and awards are pure functions over an append-only stat table, so a provider correction propagates everywhere with no cache-invalidation logic.
• Wrote 518 tests across unit, integration (Testcontainers-managed Postgres running real migrations), and component tiers, gated in GitHub Actions CI.

A retrieval-augmented chatbot that answers questions from private data instead of from the model’s own training. I built the Python backend and the RAG implementation; teammates built the ingestion pipeline and the React frontend.

• Python
• FastAPI
• Gemini
• Firestore
• sentence-transformers
• BM25

• Hybrid retrieval fusing dense sentence-transformer embeddings with BM25 keyword scoring, then reranking the top 10 candidates through a cross-encoder down to the 5 most relevant.
• Cached document embeddings by SHA-256 content hash and moved four blocking model calls onto worker threads, removing event-loop stalls under concurrent requests.
• Diagnosed a tokenizer defect that silently disabled sparse retrieval, lifting ranked corpus coverage from 6% to 100% and correcting the top-ranked result.

A modeling project to find players the draft market has mispriced, and a live decision engine that acts on it during a draft.

• Python
• scikit-learn
• scipy
• pandas
• SQLite

• Modeled a market’s price-to-value curve with per-position isotonic regression after linear, logarithmic, and polynomial fits failed to capture its shape, using residuals against the monotonic fit to identify mispriced assets.
• Built a real-time decision engine combining five normalized signals under phase-dependent weightings with Monte Carlo opportunity-cost estimation (50 seeded simulations per evaluation), run/cliff detection, and deterministic override rules.
• Reverse-engineered an undocumented private API — including a non-standard raw-JWT auth scheme and inconsistent versioned base URLs — and built a 6-stage ETL pipeline into a normalized 15-table warehouse of ~26,000 rows.

A platform for grading draft decisions after the fact — which is only meaningful if the grading refuses to use information that did not exist at the time.

• TypeScript
• React
• Vite
• Python
• pandas
• Parquet

• Eliminated look-ahead bias in retrospective decision grading by implementing point-in-time projection snapshots, ensuring every historical decision is scored only against information that existed when it was made.
• Designed a strictly pure-function evaluation core with no I/O or side effects, enabling the same ranking functions to serve both real-time forward recommendation and retroactive backward grading with zero duplicated logic.
• Reverse-engineered an undocumented commercial API with a Python probe harness enumerating ~120 endpoint permutations across two hosts and five API versions, auto-generating a live endpoint reference and a written spec.

One system, two halves. The ingestion side loads and validates multi-million-row supplier datasets; the service side exposes them through an API-driven web tier split across a network boundary.

• PHP
• MySQL
• Nginx
• AWS EC2

• Built a parallel ingestion pipeline for a 5.1M-row / 405 MB dataset, partitioning it into 52 chunks run through a hand-rolled POSIX process pool with bounded concurrency and PID-based liveness reaping.
• Eliminated roughly 10M redundant SQL round-trips by pre-loading foreign-key lookups into in-memory hash maps, and paired batched INSERTs with a row-by-row fallback on MySQL error 1062 — keeping batch throughput while retaining exact attribution for every duplicate.
• Refactored a duplicated web tier into a fully API-driven client across a network boundary, removing all database access and credentials from the public EC2 instance and replacing 14 direct-SQL functions with a cURL client layer.
• Designed a 15-endpoint REST-style API with front-controller routing (Nginx try_files → PHP-FPM), a uniform {status, message, data} envelope, and a typed data-quality layer classifying every rejected row into one of 8 error categories.