WorkSkillsRecognitionAboutContactRésumé ↓
Case Study 02 · AI Tool

Receipts

Turning raw interview transcripts into synthesis where every theme traces back to the exact quote that supports it.

Status
Live, in user testing
Type
AI tool, product
Built on
Lovable, Supabase, Claude

Where this came from

Most user research dies in the synthesis. A team runs great interviews, then someone turns hours of messy transcripts into themes, usually under deadline. The exact quotes and real pain points, the most valuable part, are the first to slip. So the PM writes users want X and no one can trace where X came from. I built Receipts to fix exactly that.

The tool is never allowed to assert a theme it cannot attach to a verbatim quote. That single rule drove everything downstream.

What it does, end to end

It ingests transcripts, infers each speaker role, and flags low-confidence inferences instead of guessing silently. It handles the unglamorous cases, same-name speakers mislabeled across files, then merges across interviews in parallel and drafts grounded PRD requirements, each one pointing at the evidence beneath it, with gaps named openly.

I measured it against ground truth

A promise like every claim traces to a real quote is only worth making if it is measured. I built an eval to test whether Receipts does what it claims, and to find where it does not. It was a record of testing it honestly, including what broke.

0
Quote fidelity, exact-substring match
0
Hallucinated or unsupported claims
0
Speaker attribution after the fix
0
Transcripts in the test set
Metric Score Basis
Quote fidelity 100% (~60 of ~60 quotes exact-substring) 6 transcripts
Hallucination rate 0% (0 fabricated quotes) 6 transcripts
Speaker attribution ~50% to 100% after a fix 4 real transcripts
Distinct-speaker integrity Pass, no voice collapsed into one 4 real transcripts
PRD grounding + gap honesty Pass all outputs
Evidence-gap calibration Finding, not a pass (below) 2 constructed cases

The finding I did not plan for

On a first run, two of four interviewees whose transcripts used generic Speaker labels were mis-attributed. I scoped a fix, deployed it, and re-ran: all four now resolve correctly, with zero drift in quote text. Then the eval caught something I did not expect. The evidence-gap flag I documented actually behaves as a filter: weak-signal themes are removed before they can become an unsupported claim, rather than surfaced and marked weak. That gap between documented and actual behavior is on the fix list.

The point was not to show the tool passing. It was to test my own product’s claims against ground truth and find where reality diverges from the pitch.

The product point underneath the tool

Synthesis that cannot be traced back to a user should not be trusted to drive a roadmap. Receipts is a bet that traceability is the valuable part, and that AI’s real leverage here is not speed, it is the discipline to keep every claim honest at a scale a human would not sustain by hand.

Next case study
FitSense