Software Engineer, AI/ML · I ship production agent systems · Cambridge, Massachusetts · B.S. Computer Science, University of Oregon
I work across the stack that production AI actually needs — relational schema design and ETL, model training and evaluation, and the agent tooling on top.
Currently: Software Engineer, AI/ML at Horizon Intelligence Labs — building production applications on Cortex, the company's AI platform, and designing the benchmarks that evaluate model capabilities.
| Languages |      |
| AI / LLM Ops |        |
| ML / Data Sci |       |
| Backend / Infra |         |
| Frontend |    |
| Role | Org | Focus |
|---|---|---|
| Software Engineer, AI/ML | Horizon Intelligence Labs | Production applications on Cortex; benchmarks evaluating model capabilities; Cortex SDK |
| Applied AI Intern | Machine & Minds | Applied AI systems built for measurable P&L impact |
| AI Engineer Intern | Modern Amenities | FastAPI + Claude sales chatbot, 900+ users, structured tool-use outputs, prompt-injection hardening |
| Lead Software Engineer | Oregon Blockchain Group | Agent orchestration on AWS Bedrock; TrialWeave (GLP-1 RWE) & Crop Share; led teams of 3–5 |
Portfolio + Justin's Bot · React Express Claude Docker Caddy
A portfolio that's also a macOS desktop simulation, a streaming Claude chatbot, and a live Coinbase order book. The bot runs a hardened, identity-pinned system prompt with prompt caching and a per-IP rate limit, and ships with a red-team suite that hammers it with 12 adversarial prompts — jailbreaks, injection via fake transcripts, base64 obfuscation, identity probes. Self-hosted on EC2 behind Caddy with automatic TLS.
Churn Prediction + LLM Retention Email Pipeline · XGBoost TreeSHAP Dagster dbt MLflow Claude Postgres
Nightly churn-to-retention pipeline scoring 18,618 customers at 0.79 ROC-AUC. TreeSHAP attributions are persisted as structured prompt input for Claude-generated retention emails, each grounded in live order and review history through read-only Postgres tools. An LLM-as-judge eval with enum-typed verdicts against explicit criteria surfaced real generator bugs — wrong CTA intent, prohibited offers, tone/label mismatch — that a rubric score alone would have hidden.
Evergreen Capital — AI-Governed On-Chain Prediction Fund · Gemini FastAPI Solana TypeScript
A 5-agent research desk (Quant, Macro, Skeptic, Data Miner, Trader) that researches independently, debates via structured cross-examination, and produces weighted consensus votes triggering autonomous Polymarket trades with no human in the loop. Full-stack MVP in 24 hours — 1st Place Solana Track, 2nd Place Polymarket Track at QuackHacks II.
Lethe — 3-LLM Consensus Medical-Bill Auditor · GPT-4o Claude Gemini 0G Chain
Three models audit hospital bills through a two-round reflection protocol — independent analysis, then peer-informed refinement — with a 2-of-3 quorum required to dispute a charge. Agents talk over a P2P mesh with ed25519 identities, so no single model can dictate the verdict. PHI is stripped before anything reaches the agents. Top 10% of 470+ projects at ETH Global.
Market Data Pipeline + AI News Synthesis · FastAPI Supabase AWS Dagster LangChain
Postgres medallion schema (Bronze → Silver → Gold → AI) keyed by (ticker, interval, timestamp) across 4.5M+ rows and 230+ tickers, with conflict-aware upserts and 100% idempotent reruns. A scheduled LangChain agent synthesizes market context, news, and filings into source-attributed summaries powering downstream RAG.