Software engineer specializing in applied AI — with the full-stack chops to ship it. RAG pipelines, tool-using agents, and fine-tuned models, built and evaluated end to end.
A production RAG service answering disaster-assistance questions with citations to official government sources, running on self-hosted infrastructure — deterministic citation validation, a published evaluation harness (1.00 grounding, 0.94 MRR), and 47 offline tests with CI.
FastAPI · PostgreSQL + pgvector · Redis · Claude (Anthropic API) · Docker
Routes each message to deliver, digest, or suppress by combining 16 deterministic signals with one constrained Claude reasoning step, plus vision OCR and local Whisper — 96.7% accuracy against a 53.3% rules-only baseline, verified by 292 offline tests.
Python · Claude (Anthropic API) · Whisper · pytest
A tool-using agent with an adaptive planning loop that branches on session state — defined tool contracts, documented failure modes, and a 77-test pytest suite with LLM calls mocked for deterministic testing.
Python · Groq · Gradio · pytest
Chunking strategy, embeddings, vector storage, grounded generation with source attribution — and the evaluation tables to prove retrieval quality, not just demo it.
Multi-tool orchestration with planning loops that branch on state, explicit tool contracts, documented failure modes, and deterministic test suites.
Hand-labeled datasets, fine-tuned transformers benchmarked against LLM baselines, confusion matrices, and honest error analysis — including where the model underperforms.