How I build AI systems that hold up in production.

Context engineering

An agent is only as good as what it is given to read. I design that context deliberately: the right records, documents and rules for each step, retrieved and trimmed rather than pasted in whole, with sources cited so every output can be traced back to where it came from. Structured data stays structured; the model extracts and drafts, and code decides anything that can be decided by a rule.

In the listing matcher demo, the language step only extracts the buyer's requirements. Budget and bedroom limits run as hard filters in code, so a listing over budget can never be presented as a match.

Harness engineering

The harness is the software around the model: the steps it may take, the tools it may call, the order they run in, and what happens when something goes wrong. I build staged pipelines with explicit inputs and outputs, schema validation on every model response, retries and fallbacks, and runs that can resume after a failure instead of starting over.

Evals on your own examples

Before go-live I test the system on real examples from your business, not on a generic benchmark. You see the results, including the failures, and you and I agree on what “good enough” means before it touches live work. The same tests run again after every change, so a fix in one place doesn't quietly break another.

Human approval gates

Anything that leaves your business, costs money or is hard to undo goes to a person first: a reply to a client, a payment, a change to a record. Approvals are designed to take seconds. As trust builds, you can decide which steps no longer need one.

Spend caps and limits

Every system has hard limits on model usage and cost per run, per day and per month, with an alert before a limit is reached. Tools get the minimum access they need: read-only where possible, and never more than the task requires.

Logging and visibility

Every run records what the system saw, what it decided, which tools it called and what it cost. When something looks wrong, there is a record to inspect rather than a guess. Failures alert a person instead of failing silently.

Tool-agnostic, and you own it

I use no-code automation tools where they fit and code where reliability or cost demands it, and I'll tell you which and why. Everything is built in your accounts, with the code, credentials and plain-English documentation handed over.

Production engineering underneath

About ten years of backend and data work: APIs and services in Python, Go and Java, PostgreSQL performance tuning, scheduled data pipelines, and cloud infrastructure on Google Cloud and AWS. Most AI projects that break, break here, in the plumbing.