Selected Projects
Full list on GitHub.
Flight recorder and firewall for AI coding agents: records every tool call, tracks lethal-trifecta taint across a session, and blocks prompt-injection exfiltration before it happens. Zero dependencies, published on PyPI, validated against an adversarial attack corpus in CI.
Scheduler and optimizer for AI coding CLIs (Claude Code, Codex, Cursor): learns your daily usage rhythm from session logs, queues work into spare capacity in isolated git worktrees, and prewarms the rolling usage window. TypeScript/Node, published on npm.
Describe your task in plain English, get one AI model recommendation with receipts — an LLM-orchestrated judge over a CI-validated dataset, with a free instant demo.
Production-grade agentic pipeline matching bank statements to invoices: rule-based matching with a Claude tool-calling fallback, 90.6% accuracy and 92.5% F1 on a 32-case golden evaluation suite, with invariant enforcement, step-level traces, and replay.
ReAct-style agent that answers financial research questions by routing between SQL over structured financials and semantic search over 10-K filings — 89% accuracy graded by an LLM-as-a-judge harness.
Translates legacy T-SQL to cloud-warehouse SQL by combining deterministic AST parsing with an LLM fallback and a self-correction loop that feeds compilation errors back until the output verifies.
Fully vectorized options backtesting engine: 10.3 million rows of 1-minute data processed in under 8 seconds with zero Python loops, synthetic pricing via put-call parity.
CLIP-based joint text-image embeddings finding near-duplicate listings in a 10,000-product catalog — 9,721 duplicate pairs at 99.94% processing success.
Real-time platform processing 10M+ patient records: Kafka + Spark Structured Streaming, dbt medallion architecture (31 models), Great Expectations data quality at 99%+, Terraform-provisioned Snowflake/GCP, Prometheus/Grafana observability.
End-to-end MLOps: injury-risk models over 4,500+ games served via FastAPI at sub-100 ms p95 and 1000+ requests/second, with MLflow versioning and Kubernetes autoscaling.
Production microservices backend: 4 FastAPI services, JWT auth, 92% test coverage, sub-50 ms PostgreSQL queries, Redis caching, 500+ concurrent users under load.
35-condition classifier trained on 245,000 images to 91.8% accuracy; training accelerated 3.37x with PyTorch DDP across 4x NVIDIA A100 GPUs and mixed precision.
Fine-tuned CodeT5 with a FAISS-backed RAG pipeline over 2.1M code-documentation pairs — 24.1% BLEU improvement over a GPT-4 baseline with sub-35 ms retrieval.
Black-Scholes and Heston calibration on implied-volatility data with convex optimization, backtested volatility strategies, and VaR/Expected Shortfall risk analysis.