Mortgage documents at Loan Labs. Bookkeeping ledgers at Finally. Learning systems research at Virginia Tech. The work is different each time; the standard is not.
Reliable AI products need strong systems beneath them. Deterministic request boundaries, type safety, fail-closed permissions. Interfaces that leave the user in charge.
Production services, typed APIs, deterministic orchestration.
Specifications, user research, evaluation harnesses.
Dense embeddings, lexical search, high-throughput pipelines.
Before production ledgers there was research. At Virginia Tech my thesis asked how an LLM can walk a student through a hard programming problem without ever handing over the answer. Guardrails were the interesting part.
Virginia Tech · 2021–2023
PICT, Pune · 2016–2020
An adaptive learning platform built on retrieval-augmented generation and Socratic guardrails, so a student gets the next hint instead of the final answer. Published at IEEE FIE 2024 and IEEE SoutheastCon 2023.
Read the research write-up →Mortgage packets run past 400 pages, and borrowers name them IMG_4901.pdf. I re-architected Lois from one-off Ruby LLM calls into a LangGraph state machine on Bedrock AgentCore: document classification, lender policy validation, and fail-closed authorization where the agent can never turn a low-confidence guess into an irreversible write.
First AI Product Engineer there. Led a three-engineer team from prototype to production on Classify AI, a hybrid retrieval engine (dense vectors for meaning, BM25 for the literal string) sorting 50,000+ bank transactions a day into per-company charts of accounts.
Bureau scores describe who a company was. We underwrote on who it is: reconstructing 90 days of daily balances, measuring cash volatility, and recalculating limits weekly with circuit breakers. $3M+ in credit across 50+ companies, zero defaults through the rollout.
Built reusable NLP inference and deployment infrastructure with PySpark, MLflow, ONNX, and NVIDIA Triton on Google Kubernetes Engine.
Software that touches financial ledgers or legal documents cannot run on prompt optimism. Four tenets I keep coming back to.
Agents need explicit state machines and typed transitions. A prompt configures behavior; a state graph guarantees you can name the step where things failed.
No agent should hold ambient API access. Reads, writes, and syncs stay scoped to verified tenants, and unknown actions get rejected at the gateway rather than apologized for in a prompt.
Vector search alone cannot match an invoice number. Meaning and literal strings need different retrievers, and the companies that get both right ship fewer wrong answers.
The best systems come from engineers who have sat in the user interview, written the schema, and watched the background worker fail at 3am. The interface is not beneath the architecture. It is part of it.
Whether you're exploring agentic architectures, evaluating financial data systems, or looking for an engineer to lead an AI initiative, I'm always glad to talk craft and architecture.