Hi, I’m Gavin. I build reliable software and performance-focused systems.

Contributed to parallel initialization and validation logic in the QuEST quantum simulation toolkit, collaborating directly with core maintainers on performance and correctness improvements.

• Worked on MPI, CUDA, and OpenMP initialization paths
• Assisted with validation utilities and performance analysis
• Received direct feedback and acknowledgment from the lead maintainer

Built a headless browser scraping pipeline for a private client to aggregate and analyze structured data across thousands of dynamic pages, handling auth flows, pagination, and anti-bot countermeasures automatically.

• Scraped and structured 50,000+ records with 99% field-level accuracy validated against source
• Reduced client's manual data entry workload by an estimated 120+ hours
• Automated nightly runs with delta detection to capture only new or changed entries

A fully local, privacy-first document Q&A system. Ask natural language questions across PDFs, Word docs, emails, and text files — LLM inference and vector storage run entirely on-device, no data ever leaves the machine.

• Hybrid retrieval combining BM25 keyword search and semantic vector search via Reciprocal Rank Fusion
• HyDE mode — LLM generates a hypothetical answer first to improve semantic chunk matching
• SHA256-based incremental indexing; only new or modified files are re-embedded on subsequent runs

A multi-agent text-to-SQL system that plans, writes, validates, executes, and explains queries from a plain-English question. Six specialized agents — schema explorer, planner, writer, validator, executor, summarizer — take turns with a self-correcting retry loop and real safety guardrails.

• Six-agent loop hand-written in ~1,900 lines of typed Python — no LangChain, no LangGraph
• Validator-driven self-correction: writer sees the prior SQL + error on retry; planner stays fixed
• Read-only execution under wall-clock timeout + row cap; safety policy enforced via sqlglot

A production-style RAG system for research papers with hybrid retrieval, a cross-encoder reranker, citation-grounded generation, and an automated LLM-as-judge eval harness that actually measures whether the answers are any good.

• BM25 + dense retrieval fused via Reciprocal Rank Fusion, then cross-encoder reranking (MiniLM)
• Citation-grounded prompt with an explicit refusal path when the top retrieval score is below threshold
• Built-in eval harness scoring faithfulness, answer relevance, context precision, and citation accuracy

An end-to-end news intelligence pipeline. Ingests articles from RSS, embeds and clusters them into topics with HDBSCAN, generates per-cluster summaries and per-article sentiment with Claude, detects accelerating topics, and surfaces everything on a 4-panel Streamlit dashboard.

• Idempotent stages — ingest dedupes on URL; embedding and sentiment skip already-processed rows
• HDBSCAN clusters headlines into topics; SQL + pandas computes today-vs-trailing-7-day acceleration
• Runs end-to-end on a single DuckDB file with no internet or API key required (sample data shipped)

A weekly AI news digest produced by spawning parallel Claude Code research subagents. One slash command kicks off the run, and the output lands as a 1–2 page blog-style "what's now possible" report.

• Fans out 4 parallel research subagents across Reddit, X/Twitter, Anthropic, and other major AI companies
• Synthesizes findings into a structured markdown digest saved per-week under reports/
• Maintains a history.md log so each run automatically skips topics already covered

A multi-source residential rental scraper that runs daily across 8+ apartment platforms (Equity, RentCafe, Bozzuto, Sightmap, RealPage, Spherexx) and writes a deduplicated, unit-level Excel report.

• Property-specific parsers handle distinct DOM layouts and embedded JSON endpoints per platform
• End-to-end audited against live pages — recovered 50+ rows per day previously lost to layout drift
• Daily delta xlsx output captures per-unit price, size, beds, baths, and availability date

A web app that scores a resume against a job description and returns structured, actionable feedback — match score, missing skills, concrete resume edits, and likely interview questions.

• Structured output via Claude API + Pydantic schemas so responses are always parseable
• PDF resume upload and parsing with pypdf, plus a plain-text fallback
• Concrete rewrite suggestions ("change X to Y") instead of generic advice like "add more detail"

A single-file 301/501 darts scoring app — server, game logic, and UI all live in one ~700-line Python file with zero external dependencies beyond the standard library.

• Browser frontend served from a Python stdlib HTTP server — no frameworks, no build step
• Server-side game state synced to the client via a small REST API
• Persistent player leaderboard saved to JSON, survives server restarts

The site you're on. Hand-rolled with no frameworks or build step — just HTML, CSS, and vanilla JS, hosted on S3.

• Drag-to-swipe project carousel with pointer events, seamless looping, and keyboard control
• Responsive layout, semantic HTML, skip-to-content link, and proper ARIA for the nav toggle
• Typography and palette tuned by hand — no UI kit, no Tailwind, no CSS framework