AI Engineer — production LLM, agentic and retrieval systems, and the verification underneath them. Boston, MA · MS Information Systems, Northeastern University (Dec 2026)
I once audited a model's confidence against its realized error and found the two uncorrelated — it was most certain exactly where it was wrong. I'd been routing human review by confidence for months.
So I build gates: deterministic verifiers that override the model when it contradicts its source, evaluation harnesses that reject changes before they ship, and review queues ranked on measured error rather than what the model believes. A wrong answer with no warning is worse than a missing one, because someone acts on it.
Shizen — sole engineer, no codebase, no team, first production release in four months. Four Python services on AWS EKS, agentic document discovery across ~33 sources, LLM extraction with page-level citation verification, and a bitemporal warehouse where every stored value traces back to its source document and date. Cost per unit went from $2.50 to $0.20; 5,457 review items cleared by folding reviewer corrections back into the pipeline as rules.
Civica — distributed PySpark ETL on Databricks for financial and healthcare clients, 50K+ documents daily, 5+ TB through Azure Data Factory, under regulated data-handling standards.
| Project | What it is | Result |
|---|---|---|
| Policy optimization from scratch | PPO with GAE and a learned critic, an advantage-weighted variant with a smooth-max reference baseline, and REINFORCE — three algorithms against Qwen2.5-3B with no reference code | 88% on countdown + multiplication, 39+ H100 runs |
| Learned agent orchestration | REINFORCE controller selecting reasoning and validation operators per problem, optimizing score minus weighted cost | 87.8% pass@1 (144/164) on HumanEval+ |
| Derivatives risk engine | Arbitrage-free SVI surfaces, VaR and ES by full revaluation and Taylor expansion, Kupiec and Christoffersen backtests, PCA factors, reverse stress | Two independent risk methods, cross-validated |
| MCTS + LLM-guided search | Monte Carlo Tree Search with UCT from scratch, extended with language-model priors for the 24-Game | 80% vs 60% heuristic baseline |
| Multi-armed bandits | ε-greedy and gradient bandits, softmax preferences with and without baseline | Sutton & Barto Fig 2.2 reproduced, 2,000 × 1,000 |
| Identity anonymization GANs | Conditional GANs stripping identity while preserving downstream-useful attributes, with representation disentanglement | Adversarial training, PyTorch |
Every figure above is traceable to a log, CSV or checkpoint in its repo. I audited these myself and removed the ones that weren't.
Languages — Python · SQL · TypeScript · Java · Shell LLM & Agents — LangChain · LangGraph · MCP · Function calling · Multi-agent orchestration · Claude, GPT, Gemini Retrieval — RAG · FAISS · Pinecone · Embeddings · Semantic chunking · Document AI / OCR ML — PyTorch · TensorFlow · Keras · scikit-learn · Hugging Face · Transformers · RL (PPO, A*-PO, REINFORCE) · Bandits · Conformal prediction Evaluation — Eval harnesses · A/B gating · Hand-labeled sets · Calibration auditing · Release gates · Failure analysis Data — PySpark · Databricks · Airflow · dbt · Azure Data Factory · Postgres · DuckDB · Redis · Snowflake Infra — Docker · Kubernetes (EKS) · FastAPI · MLflow · Terraform · CI/CD (GitHub Actions, Azure DevOps) · AWS · GCP (Vertex AI, Cloud Run) · Azure · Modal Quant — Backtesting · Temporal validation · PCA · Monte Carlo · VaR / ES · Coverage tests