AI Software Engineer building production multi-agent and RAG systems in Python and TypeScript, deployed
on AWS, Railway, and Vercel. Systems ground their outputs in retrieved evidence with citations, refuse or
hand off to a human when evidence is thin, and are gated by evals. Works spec-first with AI-assisted
development (Claude Code), pairing velocity with strict typing, infrastructure as code, and 4,000+
automated tests.
Projects
- Multi-agent memo pipeline shipped as the production default only after a paid LLM-as-judge evaluation showed it beat the single-pass baseline (groundedness 97.9% vs 95.9%, citations 100% vs 97.7%).
- A deterministic scope gate decides which laws apply, so the compliance verdict cannot be hallucinated; the model writes only cited explanation prose. Grounded RAG over 12 AI-regulation laws across 7 jurisdictions, each in its own operative legal test; adding a law is a data change, not code.
- Same engine exposed as an MCP server; a monitoring agent opens pull requests as the human gate.
- Python, FastAPI, Streamlit, Chroma, fastembed, Claude Haiku + Sonnet + Opus, GitHub Actions CI, Railway
- Agent walks a 9,000-edge music-lineage graph built from Wikidata and DBpedia. The model proposes claims, a deterministic Python gate checks each against the versioned corpus, and prose is written only from approved claims, so every connection cites its source or is refused.
- Release-gated evaluation: 63 live cases, six blocking gates (grounding, citations, prompt injection, refusal accuracy), bounds from a measured noise floor, a sealed held-out set, and a public report showing failures.
- Serverless, all in Terraform: a hand-built Bedrock Converse tool loop on a Lambda container streaming SSE, S3 + CloudFront, keyless OIDC deploys from GitHub Actions.
- Python, Bedrock (Claude Haiku 4.5, Nova Pro), AWS Lambda, S3, CloudFront, Terraform, React
- Multi-tenant RAG: each user's GEDCOM tree embeds into a private Pinecone namespace, retrieved with a public emigration corpus; a three-node LangGraph pipeline hands off to the user when retrieval confidence is low rather than inventing ancestry.
- Python FastAPI evaluation microservice scores every narrative on six metrics (Ragas, LLM-as-judge), regression-tested in CI with promptfoo and traced in LangSmith.
- TypeScript, LangGraph, Pinecone, OpenAI GPT-4o + Whisper, ElevenLabs, Python FastAPI, React 19
- 12 scout agents poll nine government data sources 24/7; a five-agent swarm on a Redis Streams pipeline correlates events against 750 IUCN species ranges in PostGIS and posts Discord alerts.
- Self-improving Refiner/Evaluator loop scored real flood and coral-bleaching predictions against observed outcomes and rewrote its own agent prompts in production.
- TypeScript, Node.js, Express 5, Redis Streams, PostGIS + pgvector, Gemini, Claude, Next.js 15, Railway
Also live: Asteroid Bonanza (four-agent swarm over 35,000+ near-Earth objects with live NASA tool calls),
Poster Pilot (multimodal RAG over 5,000+ historical posters, CLIP + Reciprocal Rank Fusion),
SoilProve (fertilizer-prescription tool, multi-provider LLM failover, built solo in 5 days), and
ChronoQuizzr (geography trivia, adversarial two-agent clue-verification pipeline).
Experience
AI Software Engineer · Self-directed · Remote · January 2026 – present
Designed, built, and operate the production AI systems above: agent orchestration, grounded RAG,
evaluation pipelines, streaming APIs, an MCP server,
cost-aware multi-model routing, and serverless AWS in Terraform.
Student Software Engineer (Apprenticeship) · Codefi · Remote · January 2025 – present
Year-long full-stack bootcamp (JavaScript, TypeScript, Angular, React, Ruby on Rails), followed
by Codefi's AI Skills track: Prompt-to-Product and the production AI engineering Masterclass.
Education
- Codefi — AI Engineering & Skills Training: Prompt-to-Product & AI Masterclass (2026)
- Codefi CodeLabs — Full-Stack Web Development Bootcamp (2025)
- SMU Dedman School of Law — J.D.
- Indiana University Bloomington — B.A.
Skills
AI / LLM: Retrieval-Augmented Generation (RAG), multi-agent orchestration, LLM-as-judge evaluation, golden datasets, Model Context Protocol (MCP), LLM APIs (Anthropic, OpenAI, Google, Amazon Bedrock), LangGraph, vector databases (Chroma, Pinecone, pgvector), prompt engineering, spend-safety
Languages & backend: Python (FastAPI, Streamlit), TypeScript (strict), Node.js, Express 5, Ruby on Rails, PostgreSQL, PostGIS, Redis Streams, SSE streaming
Frontend & testing: React, Angular, Next.js, Tailwind CSS; pytest, Vitest, Playwright
Cloud: AWS (Bedrock, Lambda, S3, CloudFront), Terraform, Docker, GitHub Actions + OIDC, Railway, Vercel