Most AI projects do not fail at the demo. They fail in the ninety days after it, when the prototype meets real users, real documents, real permissions and a real bill. The gap between “the model can do this” and “this runs in production” is engineering work, and it is specific: control flow, retrieval quality, evaluation, guardrails, cost, latency, observability, rollback.
That gap is what I am hired to close.
What you are hiring
- Agent systems that terminate. Planning loops with explicit budgets, tool contracts, retries and cutoffs - so a bad step costs you a few seconds, not a runaway bill or a corrupted record.
- Retrieval that respects your org chart. Permission-aware RAG, hybrid search, chunking that matches your document reality, and freshness you can reason about.
- Evals before opinions. A golden set, a regression gate in CI, and a number that tells you whether last night’s prompt change made things better or worse.
- Guardrails and red-teaming. Prompt injection, tool abuse, data exfiltration and jailbreaks are treated as a threat model, not a disclaimer. I maintain an open-source LLM red-teaming kit used by other teams to test their own systems.
- AgentOps. Traces, replay, cost per task, failure taxonomies, and a rollback path that does not require a heroic night.
- A team that can take over. Handover is the deliverable, not an afterthought. Your engineers should be able to extend the system without me in the room.
What you are not hiring
Not a wrapper vendor. Not a platform with seats to license. Not a research lab that returns a paper. Not a body shop that staffs four juniors against a statement of work and calls the overhead “velocity.”
One senior engineer, working directly with your team, on a scope small enough to finish and specific enough to verify.
How engagements usually run
- Twenty-minute call. What you are trying to build, what is already in place, what “done” means. Free, and you get a straight answer even if the answer is “you do not need this.”
- Discovery Sprint - 2 to 4 weeks. Architecture, a working thin slice, a security pre-check, cost and latency modelling, and a roadmap your board can read. Fixed fee, credited against the build.
- Build - 4 to 10 weeks. Embedded with your team, shipping behind flags, with evals gating every release.
- Handover or fractional. Either your team owns it outright, or I stay on a recurring block for the next thing.
Full detail on scope, pricing structure and commitment lengths is on the rates and engagement models page.