Evidence-based AI workflows
Make the output useful enough to trust.
An AI workflow needs a clear job, evidence behind its output, and a way for a person to check the result. I help separate reliable rule-based checks from model judgment and define how the workflow will be evaluated.
Work directly with Hasan Burak Taşyürek at HB Global.
Scope, fee, timing, and payment terms are agreed before paid work.
What the work gives you.
The agreed scope determines the depth of each output.
- A defined review task
- The decision the workflow supports, the evidence it may use, and the points where a person retains responsibility.
- Checks with inspectable evidence
- Rule-based checks where possible, with model judgments tied to the source material and exposed for review.
- An evaluation and fallback plan
- Agreed examples, review criteria, and failure behavior so usefulness can be examined on a defined sample rather than assumed from a convincing demo.
How I approach it.
I start with the human task and the criteria for a useful result. Rules that can be expressed in code stay deterministic. Model output is checked against source evidence before it reaches the reviewer. Evaluation records which sample was used, what was counted, and what remains outside the test. Implementation scope follows those decisions.
Where the scope ends.
A result on one annotated sample is not a general accuracy guarantee. Evaluation coverage, available references, and the role of human judgment stay explicit. A model response does not replace the reviewer’s responsibility for the final decision.
The work behind the approach.
Each case describes its own context, decisions, and limits.
An AI-assisted PowerPoint review workflow
On the annotated sample deck, findings fell from 110 to 31 while retaining 20 of 22 detectable reviewer-noted issues. The case explains the evidence gates, review criteria, and two-deck delivery scope.
Read the caseBefore we start.
Does every check need AI?
No. A rule that can be checked reliably in code usually belongs there. Model judgment is useful where a deterministic check would be too brittle, provided its output can still be reviewed.
How do we decide whether the workflow is better?
Agree a representative sample and review criteria first. Examine missed findings, noise, and whether the surviving output helps a person make the intended decision. Keep the limits of the sample visible.
Can we begin with an existing prototype?
Yes. Describe what it should help someone do, where trust breaks down, and how a reviewer checks the output today. That is enough to begin a scope discussion.
Start with the decision.
Describe the system, the constraint, and what a useful outcome would look like.
Discuss your situation