Work with me
Building AI that has to work in production.
Here's how I work with teams building AI agents and RAG systems. Not a menu of packages, but a description of the shapes this work usually takes, and how I think about it.
Where I'm useful
“Our team wastes hours searching scattered internal docs.”
A retrieval problem dressed as a search problem. Usually the fix is in the retrieval quality and the data, not a bigger model.
“We built an agent POC and it doesn't hold up.”
The demo worked; production doesn't. The gap is evaluation, failure handling, and the architecture around the model, the parts tutorials skip.
“We don't know if this AI project is worth doing.”
Before you build, it's worth an honest read on the ROI, the risk, and whether an agent is even the right shape for the problem.
How that usually looks
A workshop or working session
A focused session with your team to pressure-test an idea, an architecture, or a roadmap. Best when you want to align on direction before committing engineering time.
A short technical review
A read of an existing agent or RAG system: where it will break, what to evaluate, what to change first. A few days, ending in a written assessment you can act on.
A scoped pilot build
Building a narrow, real slice end-to-end, with the evaluation and production concerns in from the start, so you learn whether the full thing is worth it.
How I think about this work
Start with a workflow, not an agent
Most problems that get reached for an agent are better served by a constrained workflow. Reach for autonomy only where it earns its keep.
Evaluate before you scale
Without evaluation you're shipping vibes. A system you can't measure is a system you can't improve or trust.
RAG quality is a retrieval problem
When answers are wrong, the model is rarely the cause. The retrieval, the chunking, and the data are where quality is won or lost.
Production is everything around the model
The prompt is the easy part. The system (failure handling, observability, boundaries) is what makes it production instead of a demo.
These come straight from the writing, where the arguments are made in full.
This isn't a pitch. If any of the above sounds like where you are, the best next step is a conversation. Get in touch.