Senior AI Platform Engineer — 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.

Nine years of production engineering, the last several on LLM and agent platforms.

A multi-tenant conversation-intelligence platform turning ~100 hours of calls a week into structured business intelligence — nine queue-driven workers across seven queues each with its own dead-letter queue, retrieval over 3,072-dimension embeddings in pgvector, a 24-tool function-calling assistant, and an MCP server exposing the intelligence layer to external agents. Tenant isolation by Postgres row-level security in a single deployment. Compute cost down ~94% by moving off per-task serverless containers to capacity providers tuned per workload shape, autoscaled on queue backlog per task rather than CPU.

A 10-agent NLP pipeline on a decentralized data platform — a meta-orchestrator fanning out to seven stateless model agents in parallel, merging through a single writer into a social graph. Content-hash idempotency made the whole thing replayable: wipe Postgres, rebuild from source events, get identical output. Pipelines declared as infrastructure-as-code, so a new one ships as a stack file without recompiling any service.

An embedded AI companion platform — custom ESP32-S3 hardware, four memory layers across Redis, Postgres and a graph store, and a streaming voice loop under 500ms. Every session reprocesses the full prior history rather than reading a fixed window, so recall doesn't degrade as the window slides.

Earlier: founding engineer at a legal-AI startup through SOC 2 Type 2, and ML platform tooling in biotech.

Keeping agents accountable while they build — the problem I keep returning to. Agents that write code are easy; agents you can trust unattended are not.

• loop-kit — a portable autonomous build loop. Fresh context per iteration, builder separated from judge, and evidence-gated marking so a feature can't go green unless it re-proves green.
• sprint-harness — Shape Up + SPARC with drift control, using inject-violation-catch-restore to verify the guardrails actually fire.
• inspector-hook — cross-IDE visibility into what AI coding assistants are doing to your files, in real time.

Agentic systems

• jobsearch-automation — a local-first automation OS: ingest, score, tailor, then a hard human gate before anything is sent. Next.js dashboard over Postgres, with an LLM harness that runs faithfulness checks against fabrication.
• Topic-Modelling — real-time chat analysis into a social graph with relationship intelligence and LLM-driven topic modelling.

Architecture

• systems-architecture-studies — C4, sequence, and component diagrams for two production systems, written as pattern studies. Autoscaling on backlog-per-task, fan-out reads with a single writer, a promotion ladder that never demotes, content-hash idempotency.

📍 New York, NY · Remote (US)   ·   🔎 Open to Senior AI Platform / AI Infrastructure roles

Website · Résumé · LinkedIn · Email