AI engineering
I build with existing models: agent workflows, tool integrations and evals. I do not train models. Most of this runs in daily team work. The rest are personal projects.
SELECT * FROM work WHERE topic = 'agents'
Coding agents in a team
At work I built a workflow that takes a change from request to merge request. Most of the team uses it.
Independent review
Two agents review the diff separately: one for domain correctness, one for operational reliability. Their findings stay drafts until a person publishes or rejects them.
Shared skills and MCP
Skills and MCP integrations for development, incident investigation and code review are shared, so the whole team works the same way.
Review of the result, not the diff
For the human side I built verifair. Agents publish change reports with their evidence, and people sign, comment and approve them.
A company-wide agent kit
My own initiative. I designed and built a shared kit of agent skills and rules, an onboarding site generated from the same files, and a sandbox where an agent takes a task from chat. A working version, proposed for company-wide use.
Testing an agent platform
I ran real, low-risk tickets through Paperclip, an agent orchestration platform, reported what broke, and contributed skills and fixes to the company setup.
SELECT subject, result FROM evals
Evals
Judges for API checks
Smoke checks for internal APIs use a Pydantic AI judge that reads live responses. An eval suite tests the judges on known-good and known-bad cases.
Agent instructions
Each case runs several times, with and without a skill. One workflow skill scored 0.87 against 0.40 without it. Notes →
One more page
In ai-newsletter, a daily AI digest that checks its sources, a model may read extra pages. It scored 0.92 against 0.70 for a single pass on keeping the key condition. Notes →
Braintrust Python SDK
Reported two bugs in its Pydantic AI tracing while running evals. Both were fixed upstream.
SELECT * FROM personal_projects WHERE topic = 'ai'
Personal tools
Personal assistant
in useAn assistant I use myself, with long-term memory, tools and recurring tasks: email drafts, source-linked digests, voice queries. Built from open-source parts such as Hermes Agent, Hindsight and Home Assistant.
Proof of concept of an AI shell: a fast model answers and hands work to background Claude agents. Risky steps wait for approval.
SELECT lesson FROM experience
Principles
- An agent saying “done” is a claim, not a result. Tests and checks decide.
- Limited permissions, and human review for sensitive actions.
- For simple or critical operations, plain code is often the better choice.
- AI helps most with messy input: reading, summarizing, finding context, drafting.