- Ship AI products end-to-end: data pipelines → RAG/LLMs → UX → deployment
- Build predictive models for forecasting, churn/LTV, risk scoring, anomaly detection
- Blend Python, TypeScript, dbt, BigQuery, GCP/AWS, Docker to deliver reliable, observable systems
- Love LLM evals, prompt tooling, vector search, analytics, and automation
Right now
- 🔭 Building AI compliance tooling and data apps
- 💬 Ask me about React/TS, Node, Python, data pipelines, RAG, OpenAI, LangChain
Project highlights (click to expand)
- Benchmarks SARIMA, Neural Prophet, LSTM, RF, MLP with GPT-4 vs Claude-3 in autonomous vs assistive modes
- Representative result: Neural Prophet RMSLE ≈ 0.1458 (GPT-4, role: Data Scientist); explores prompt-role/sentiment effects
- Repro tips: deterministic seeds, leakage guards, schema validation
- Serverless RAG over 150k+ games + rulebook PDFs
- Kendra retrieval + Bedrock (Claude), optional OpenSearch vectors, Cognito auth, Amplify UI
- Glue/Step Functions ingest; CI/CD via CodeBuild/CodePipeline; cost/monitoring guardrails
- Spam detectors across email/SMS/YouTube using TF-IDF+NB, FastText+MLP, DistilBERT
- In-domain: DistilBERT up to 0.992 acc; under domain shift all models drop—analysis of precision/recall trade-offs
- Future: domain adaptation, hard-negative mining, lightweight adapters
- LendingClub 1.3M rows; application-time only (no leakage)
- ROC AUC ≈ 0.73 with Random Forest/XGBoost; engineered features (FICO composite, credit history length, installment/income)
- Next: calibrated probabilities, cost-aware thresholding, LightGBM/CatBoost
I enjoy shipping useful tools for data-heavy teams. If you’re exploring AI copilots, RAG systems, analytics platforms, or AI automation, ping me — happy to chat.
“Ship small, measure, iterate.”