IKIdrees KamalLet's talk
Selected work

INDEPENDENT PROJECT / APPLIED AI

Surface the blocker before the handoff.

A restaurant launch-readiness prototype that keeps findings tied to source evidence.

Reference prototypeNext.js · OpenAI Responses API · Structured output · Citation validation · Evaluation harness

The problem

A handoff can look complete while source documents disagree, approvals are missing, or evidence belongs to another location.

My contribution

I selected the problem, built the prototype with Codex, defined structured output and citation checks, and created deterministic safeguards and reference scenarios.

Decisions that mattered

  1. Represent readiness as individual checks with supporting quotes, not an unqualified generated summary.
  2. Downgrade a finding when its citation does not match the supplied source text.
  3. Use fictional restaurant data and require human review before a launch decision.

What changed

Built a reference walkthrough with six readiness checks, conflict and missing-evidence scenarios, structured output, and six passing deterministic safeguards in the September 4 review.

INTERACTIVE REFERENCE · FICTIONAL DATA · NO AI CALL

One closing time.
Two conflicting sources.

Sales handoff“Table & Thyme closes at 9 p.m. on Fridays.”
Menu approval note“Friday ordering runs until 10 p.m.”
Reference finding: human clarification needed

Both sources concern the same location but disagree. Do not infer a closing time or call the launch ready.

These transitions are hand-authored to explain the design. They are not model-generated results or proof of model accuracy.

Discuss this work