AI / Full-Stack Engineer at Enidus · NYC — open to relocation
I build production LLM systems, agentic copilots, RAG, and the eval harnesses that keep them honest. B.S. Computer Science + B.S. Data Science, UW–Madison 2025. My bias is that an agent claim isn't real until there's a number behind it and a test that can fail.
Building an agentic copilot for T-Mobile for Business, in pilot with 15 reseller tenants representing 25+ enterprise customers and 100+ daily portal users. It answers natural-language queries over telecom account data and runs multi-step account actions such as device purchase, line suspension, and plan upgrades.
- 53 intents dispatching to 43 Pydantic-typed tool handlers, with the LLM constrained to tool selection — never raw SQL.
- Every write-capable transaction is staged for human confirmation before the backend executes it.
- A parametrized pytest suite (52 cases fanning out to 400+ invocations) that caught the agent inventing device SKUs and malformed account numbers early.
Open to AI Engineer roles on small AI-first teams shipping real production LLM systems.
CloudGuard — a reliability and safety harness for LLM cloud agents, measured against a real AWS environment. An embeddings-based tool router recovers selection accuracy to 1.00 where a bag-of-words router degrades to 0.83, with every headline number written to a committed JSON artifact.
GoodEnough — a pre-registered study of when a 1.7B local model can replace a hosted 70B. Across a pinned case study, zero of eight benchmark slices cleared a non-inferiority margin fixed before any data was observed, and hosted answered 8.5x faster at the pooled median — the local path lost on accuracy and latency at once.
ClaudeJob — an agentic pipeline that ingests live job postings, tailors a structured-output JSON resume per role, and renders pixel-matching PDFs. Guarded by 149 unit tests and a validator suite that catches 30+ AI-resume cliché patterns and fabricated stats against a pinned base.
AI/LLM — Claude & OpenAI APIs, tool calling, agent orchestration, RAG, vector search (Qdrant), structured outputs (Pydantic), streaming/SSE, MCP, PyTorch.
Stack — Python, TypeScript, SQL, FastAPI, Node.js, React, Next.js, PostgreSQL, Docker, AWS.