Grounded QA agent over 150 arXiv ML papers — my hand-rolled ReAct agent rebuilt as a typed LangGraph state graph with checkpoints and fail-closed guardrails. A frozen 69-item eval set drives a regression gate; the LLM judge is checked against 25 human-scored answers. Public FastAPI service on Hugging Face Spaces with SSE streaming, per-IP rate limiting, a restart-durable $0.50/day cost ceiling, and Langfuse tracing. Billing reconciliation showed only 24% of 27.27M billed prompt tokens had been recorded — the post-mortem publishes that, and where v3 scores below v2.1.

LangGraph Gemini FastAPI Langfuse Docker HF Spaces

Two-stage VLM agent for insurance-claim adjudication — perception is split from reconciliation, so claim text and EXIF metadata cannot move a verdict. Ranked 16 / 1,773 — Top 1% at HackerRank Orchestrate (June 2026). 95% valid-image accuracy on 20 labeled claims (29 images), up from 60% single-pass; verdicts held on 7 of 8 adversarial injection twins; est. ~$0.017/claim (44 claims, cold cache).

Claude Sonnet 4.6 VLM Prompt Engineering Eval-Driven Dev

Personalized notify / digest / mute routing for 110 text, image, and voice messages. The model perceives, the code decides: one Claude Sonnet 4.6 call returns a schema-forced reading, and a deterministic 4-layer precedence ladder sets the action, reason, and confidence. Ranked #70 / 1,983 at HackerRank Orchestrate (August 2026). The eval report fixes its noise band before measuring and keeps the result that cut against the design: with the safety layers removed, accuracy was 27/30 vs. 25/30 — within noise.

Claude Sonnet 4.6 Vision faster-whisper Deterministic Rules Ablation

LLM medication-logging agent with a deterministic safety spine — the model is load-bearing at exactly two edges, everything between is deterministic Python. Writes are gated by a single-use signed JWT (unique JTI, 300s TTL, subject binding), so a compromised model provably cannot log an off-schedule dose. A pass^k (k=5) harness surfaced 2 reflected injection leaks a single run missed.

MCP Gemini 3.1 Pro HMAC + JWT SQLite HITL

5-agent Microsoft Agent Framework system on Azure OpenAI and Microsoft Foundry — Planner-Executor plus a Critic/Verifier loop-back (capped at 2 iterations) that can flip Not Ready → Ready once it adds study hours. Hybrid BM25 + TF-IDF retrieval fused with a Microsoft Learn MCP client, all three Microsoft IQ layers, and an MLflow eval harness. Synthetic data only: 12 learners, 3 teams.

Microsoft Agent Framework MCP Azure OpenAI RAG Streamlit

End-to-end MLOps system on 284,807 transactions (0.172% fraud) — GitHub Actions CI/CD with a training workflow, MLflow Model Registry (PostgreSQL on Amazon RDS, artifacts in S3), and a Dockerized FastAPI service on AWS EC2 behind Nginx with /predict + /predict-batch, API-key auth, and rate limiting. SMOTE applied to the training split only.

MLflow FastAPI Docker AWS GitHub Actions XGBoost

SARIMAX vs. Prophet vs. XGBoost on 3,300+ records with rolling cross-validation. Seasonal SARIMAX selected at holdout MAE 138.14 vs. CV MAE 137.86. 30-day forecasts with 95% confidence intervals in a Plotly Dash dashboard on Render.

SARIMAX Prophet XGBoost Plotly Dash Docker

BERT + Wav2Vec2 + ViT cross-modal Transformer fusion (4 layers, 8 heads). 59.96% accuracy, 60.66% weighted F1 on 7-class MELD, with class-weighted loss for label imbalance.

PyTorch BERT Wav2Vec2 ViT HuggingFace