About
Nine years, mostly infrastructure
I build the infrastructure that agent systems run on — orchestration, retrieval, evaluation, and the cost and latency controls that decide whether a system survives contact with production.
That sentence is doing real work. The interesting problems in this field are mostly not about models. They are about what happens when a pipeline fails at stage six, when ten agents want to write to the same graph, when the bill arrives, and when someone asks whether the thing you built can be re-run.
What I’ve been doing lately
A multi-tenant conversation-intelligence platform — nine queue-driven workers turning about a hundred hours of calls a week into structured intelligence, with retrieval over 3,072-dimension embeddings and an MCP server exposing the whole layer to external agents. Before that, a ten-agent NLP pipeline where seven stateless model agents fan out in parallel and a single writer merges into a social graph — content-hash idempotency made the whole thing replayable from source events.
Alongside both, my own company: an embedded AI companion platform built primarily in Rust, with custom ESP32-S3 hardware, three memory tiers — Redis, a vector store, and a knowledge graph, queried in parallel every turn — and a streaming voice loop tuned to a 520ms target against an 800ms hard budget.
Before that
Founding engineer at a legal-AI startup, building the platform from zero and taking the company through SOC 2 Type 2. ML platform tooling in biotech — a model-federation dashboard that removed most of the manual coordination from scheduling training jobs, and the training-node topology rendered as an interactive graph. Earlier still, the ordinary and useful work: monolith-to-microservices migrations, React frontends, serverless services.
How I work
I care disproportionately about the properties a system has when nobody is watching it. Is it replayable. Does the state machine have a path backwards. Is the scaling signal honest or merely available. Those questions are unglamorous and they are the difference between a system you can operate and one you can only demo.
I also think most architecture writing is too polite about tradeoffs. Every decision in the case studies has a cost, and I’ve tried to name it rather than present a list of wins.
Where I am
New York, NY, and open to remote across the US. Currently looking for Senior AI Platform or AI Infrastructure roles.