Build it right
“The demo works. Now it has to ship.”
What an agent is made of, stage by stage: retrieval, the model and, on a phone line, speech in and out.
See one call, stage by stage
Hands-on evaluation engineering for AI agents.
I help CTOs and engineering leads evaluate customer-facing LLM agents, starting with established software products adding agents.
11+ years shipping software. Production AI since 2021. MS CS (ML), Georgia Tech, completing 2026.
In a 30-minute discovery call we cover what you're building, what's failing and which kind of help fits.
Every AI change gets inspected before it reaches your users.
Three places teams get stuck
Each opens a working example you can explore.
Know it works
“You can't measure the vendor's demo.”
When a prompt, model or provider version changes, test real tasks for regressions before customers find them.
Break the demo, watch the gateKeep it working
“A wrong answer reached a customer.”
Every stage traced in production, and scripted conversations that check latency and correctness against a spec, so you see which stage regressed.
See a regression traced to its stageRecent work
Evals for AI used by scientists
At BenchSci I owned evaluation end to end. Manual validation time fell 60%.
Read the research-assistant case studyIndependent AI/ML engineering
Independent contractor (Cohen AI Consulting LLC), AI/ML engineering for a US healthcare-technology company, Jan–Oct 2026.
See one call, stage by stageHow we can work together
Two ways: an embedded AI engineer who ships features with the evaluation pipeline that measures them, or a written review of your AI system's failure modes and eval gaps. How engagements work
Start with a discovery call
Tell me what you're building, what's failing and when you need help. We'll discuss whether an engineering engagement or a written review fits.
Book a 30-minute discovery call