AI engineering & evaluation
When an AI workflow gets stuck, takes the wrong action or breaks after an update, you need a change you can verify.
The working agreement
One workflow, a clear boundary and a named owner. Decide what a good result looks like before work starts.
Work in your repository. Keep the important decisions, tests and remaining questions visible.
Review the result against the agreed cases, with a practical handover and a defined support window.
Ways to work together
Fix one reproducible failure in an existing AI workflow. Keep the regression check that catches it next time.
Find the important failure modes, test the controls and decide what to fix before committing to a build.
Build one operational AI workflow with evaluation checks, human approvals and a handover your team can run.
Calibrate an existing evaluation workflow, add a release gate and keep an evidence-based scorecard as the system changes.
Each engagement stands on its own. Timing and acceptance criteria are agreed before kickoff. Applicable taxes and payment details are confirmed in the proposal.
Selected engineering work
Public repository / Evaluation
Deterministic checks for citations, injection patterns, personal-data patterns and lexical support, with synthetic fixtures and mutation checks.
These checks make no model calls. They do not establish the accuracy of an LLM judge.
Read the codeTechnical case study / Agent systems
A documented approach to recall, handoffs and provenance, with a local verification record and a clear production boundary.
Local verification is documented. Production identity and credentialed cutover were pending at the published checkpoint.
Read the case studyBefore we start
Clear terms make it easier to decide if the work fits.
Yes, if one failure can be reproduced in an existing workflow and the fix can be bounded. If the cause is unclear or the workflow needs a broader review, start with an audit or a separately scoped investigation.
We agree the expected behavior, the test cases and the person who accepts the result before work starts. The handover includes the change, the checks and the remaining limitations.
Work is delivered into your repository. The written agreement defines your rights to project deliverables, any pre-existing tools and third-party licenses. Use synthetic or fully redacted examples for scoping; access and data handling are agreed before delivery.
No. Each offer stands on its own. A qualifying $4,000 audit is fully credited toward a $15,000 sprint for the same workflow, leaving $11,000 to pay. The evaluation retainer can also support an existing system.
Bring the workflow
Start with a 30-minute call. We’ll check whether there’s a useful, bounded piece of work to do together.
