Production-grade
AI agents,
built and measured.
I build AI agents that retrieve, reason and take action, and the evaluation layer that gates them: golden sets, regression tests in CI, tracing, guardrails. Built in your repo.
An example run: the request is planned, context is retrieved, a tool is called, a critic checks the result against sources and policy, and the run passes the eval gate.
- production AI, ML & data
- 7+ yrs
- measure, then architect
- eval-first
- no handover cliff
- your repo
- aws · mlops
- python · go · ts
What I take on
AI agent build →
Agents that retrieve, call tools and take action: retrieval and planning, multi-agent orchestration, tool and API integrations, human hand-offs. Delivered as code your engineers own.
Evaluation & guardrails →
Golden datasets, offline and online eval, regression testing in CI, tracing and observability, safety and policy guardrails. You stop shipping on vibes.
AI development partner →
Ongoing: architecture review, hands-on delivery alongside your team, hiring input, and the unglamorous work of making a system observable and operable.
Hi, I'm Rene.
I've spent 7+ years shipping production AI, ML and data systems in fast-moving startups: a multi-agent fraud detection system at Midnite, the evaluation framework and a $350K MRR AI product at Spotlight, and Perlego's first data platform as an early data hire.
Based in San Francisco, working remotely across EU and US hours.
Works with the stack you already run
- Pythonml · fastapi
- TypeScriptweb · tooling
- Goservices
- Terraformaws · infra
- LangGraphorchestration
- pgvectorvector search
- OpenTelemetrytracing
- MLflowexperiments
- Your repo, your CI.
- No vendor lock-in.
- Self-hosted models welcome.
How it fits together
Sources & index
Documents, threads and records are chunked, embedded and tagged with metadata, so retrieval can filter before it ranks.
Planner & specialist agents
A planner breaks the request down and routes each part to a retriever, a tool or a critic, retrying when a step comes back thin.
Traceable answers and actions
Every claim points at the span it came from, and every action at the policy that allowed it.
Guardrails & eval gate
Policy checks and escalation paths wrap the whole flow, and a golden set in CI blocks any change that makes the agent worse.
If it isn't measured, it doesn't ship.
- Golden set from real traffic. Built from the questions your users actually ask, not the ones that demo well.
- Gates every pull request. Prompt, model, retrieval and tool changes are scored in CI; a regression blocks the merge.
- Traces you can read. When an answer goes wrong, you open the run and see which step failed.
Notes from the field
A Playbook on Considerations when building AI Applications in Production
Why most AI demos die in production — and the framework that gets them to survive
Write-Audit-Publish (WAP) pattern for robust data pipelines
This is a guide on how to implement a WAP pattern in snowflake and dbt
How we cut our Snowflake bill in half overnight
Controlling Snowflake Costs: Recommendations and Common Pitfalls to Avoid
Tell me what's breaking.
A few lines is enough. I reply within two working days, and if it isn't a fit I'll say so and point you somewhere better.
Message sent.
Thanks for reaching out. I reply within two working days, usually sooner.