AI systems that do real work in production
The demo is not the hard part any more. The hard part is an agent that runs on a schedule, costs a predictable amount, can be audited after it does something surprising, and does not repeat an action because a step was retried. Retrieval is already serving a live consumer product of mine; the agent framework runs in beta.
Sold on the auditing, cost control, idempotency and evaluation — the parts that decide whether it survives contact with production.
Where it actually goes wrong
A model call is the easiest component in the system. Everything around it is where production risk lives: what happens when the same job runs twice, how you find out what an agent actually did last Tuesday, what stops a retry loop from spending a month's budget in an afternoon, and how you know a change to a prompt made things better rather than differently wrong.
That is ordinary distributed-systems work wearing a new hat — which is exactly why it suits an integration practice rather than a prompt shop.
What you get
- An audit trail over agent runs
- What ran, when, with what inputs, and what it changed. Retrievable months later.
- Idempotent actions
- So a retried step does not send the second email, create the second record, or charge the second time.
- Cost control
- Budgets and ceilings treated as a system property, not a monthly surprise.
- Evaluation you can act on
- A way to tell whether a change improved the output, rather than shipping on the strength of a good demo.
Where I have done this
- RAG-based search and analytics running in production in What's The Deal — a live consumer product, not a demo.
- A first-party multi-agent framework with scheduled execution, auditing and a chat surface.
- pgvector semantic search over a live listings corpus in KeystoneOS.
- A vector recommender in production behind What's The Deal, fed by GCP Pub/Sub event ingestion.
- Multi-agent orchestration with scheduled execution
- Audit trails over agent runs
- pgvector semantic search
- Vector recommenders
- Evaluation and cost control
Describe it in a paragraph.
Tell me what the system is, what it is doing wrong, and what has to be true by when. I will tell you whether I am the right person for it — including when I am not.
The other three
- Integration & data migration
The promise is not that I will connect your systems. It is that your integration survives a replay, a spike, and an outage.
- Platform builds, zero to production
Built and shipped, not prototyped — auth, billing, tenancy, queues and observability included.
- Technical due diligence & rescue
A fixed-scope audit of a stalled or inherited system, delivered as a prioritized roadmap you can act on.