What I Offer
AI Agent Harnesses
Custom agent architectures using tools like Cursor, Claude Code, and the Agent SDK. Designed for real-world productivity, not demos.
RAG Systems
Chunking, embeddings, and citation hygiene so answers point back to the file that said it — not a confabulation dressed as confidence.
Automation Pipelines
Cron plus queue plus dead-letter, or GitHub Actions glue — whatever stops humans from copy-pasting between tabs.
Prompt Engineering
System cards, tool schemas, refusal boundaries — the unsexy scaffolding that keeps agents from improvising invoices.
Consultancy
Vendor demos decoded: what plugs into your SSO, what needs a GPU budget, what is a press release.
Integration
Connecting AI capabilities with existing systems. APIs, webhooks, event-driven architectures, and data pipelines.
Approach
Name the failure mode
Is the pain manual data entry, hallucinated support answers, or “we bought seats and nobody uses them”? Different failures get different tools.
Measure the boring numbers
Latency to first token, weekly tool errors, pounds per thousand rows — pick one spine metric so we know whether a change earned its deploy.
Cut scope like you mean it
First release should embarrass you slightly in features, not in reliability. Expand when operators ask for more, not when the slide deck does.
Automation or agents
What's eating time, how often it runs, any hard deadline, and your budget or range? Let's find out if we're the right fit together!