I'm a Technical Product Manager focused on AI/ML products — RAG systems, agentic workflows, and data products. I ship end to end: framing the problem, choosing the model, building the evals and retrieval layer, wiring the backend, and shipping the interface the user actually sees. Everything below is built from scratch to solve a real problem, not a tutorial rebuild.
Currently targeting Senior/Principal PM and Amazon PMT (L5/L6) roles — domain depth in consumer lending/credit risk, fraud/AML/KYC, model governance, and supply chain.
- Spec before code — every project starts as a written PRD/spec, not a prompt.
- Eval-driven — define what "correct" means and measure against it before trusting model output (gold-sample validation, Spearman correlation checks against human judgment).
- Deterministic gates over vibes — rules-based checks wherever a rule can do the job; LLM judgment reserved for what actually needs it.
- Structured logging by default — every agent/pipeline ships with a run logger, not print statements.
Each repo README has the problem, the approach, and the outcome.
| Project | What it does | Stack |
|---|---|---|
| Credit-Risk-Monitor | Champion/challenger credit-risk model (Logistic Regression vs. XGBoost) on real 2022 HMDA Texas mortgage data — 868K+ applications, fair-lending parity screen, PSI drift monitoring, interactive Texas county risk heatmap. Live demo → | Python · XGBoost · Streamlit · HMDA Data |
| Lending-Policy-Simulator | Champion/challenger consumer-lending policy simulator on real Lending Club data — RAROC-based economic analysis (not just accuracy), vintage loss-curve analysis, fair-lending screen. Documents a case where a statistically better model didn't translate to better economics. Live demo → | Python · scikit-learn · Streamlit · Lending Club Data |
| Sediment | Read-only analytics tool for Claude Code usage — stratigraphic HTML dashboard over your own session transcripts. Zero dependencies, 61 tests, all correctness defects caught against real transcript data. | Node.js · Zero-dependency CLI |
| TrueCut (WIP — repo coming soon) | Agent cost-and-value control tool — offline/replay architecture with a deterministic benchmark suite so you can evaluate agent changes without burning live API spend. Apache-2.0. | Python · Eval harness |
| Resume Optimizer / PMT Scorer | Deterministic scoring layer + LLM judge for resume-to-role fit, calibrated against a gold set (Spearman ρ = 0.857). Identified quantification as the highest-leverage gap in bullet writing. | Claude API · Prompt Engineering · Eval harness |
AI / ML & LLMs
Languages & Data
Cloud & Infra
Product & Analytics
🌱 Currently going deeper on: vector databases (pgvector / Pinecone), LLM observability (LangSmith), and multi-agent orchestration.
- The economics of a "better" model — why a statistically stronger challenger model lost on RAROC, and what that means for how PMs should evaluate model upgrades.
Building AI products that ship. Open to Senior/Principal PM and Amazon PMT conversations.