Problem
The Mom Test is simple on paper — ask about the past, not the future; don't pitch; don't trust compliments — and nearly impossible to apply live. In a real discovery call the founder asks "would you find this useful?" and collects a fake yes. They explain the product instead of listening. They hear "great idea" and file it as signal. They accept "yeah, I'd use that" without ever digging into what the person actually did last time the problem hit.
Every one of those is a split-second mistake that is invisible while it happens. The call ends, the founder thinks it went well, and the transcript is full of false positives. Wrong validation, wrong product, dead startup.
Why I built it
Because nothing on the market coaches the quality of the conversation. Fireflies, Otter, and Granola are general meeting bots — they summarize and extract action items, but they know no methodology, suggest no questions, and never tell you that you just pitched. The book is static. A mentor who sits in on every call doesn't scale.
So the gap was specific and narrow: the discipline, applied live, by something that is already listening. Transcription is a commodity I can buy. The defensible part is the engine on top of it.
Transcription is everyone's. We sell the discipline.
Solution
A note-taker bot joins the Zoom, Meet, or Teams call, states consent, and streams the transcript. The other side never sees any AI — all of the intelligence lives in a second window only the interviewer sees.
That dashboard is deliberately quiet. One suggested question on screen at a time, plus silent warning badges — you just pitched, that's a leading question, that's a compliment, not evidence. Anything more would cost eye contact, which is the opposite of what the method is for.
Every product idea is a vault: you open it by writing down the assumptions you intend to test — problem, behavior, and value — not by describing the idea. Every live suggestion and every report traces back to one of them.
Two reports, two readers
The call ends and produces two separate documents, which is the part I care most about:
- The insight report — what the user's pain actually is, in their own words, with each vault assumption marked validated, contradicted, partial, or open. Every finding carries a signal-quality label saying whether it rests on real behavior or on a hypothetical.
- The coaching report — where you went wrong. Talk-time split, leading questions and the moments they happened, points where you pitched, compliments you mistook for evidence, follow-ups you left undug, and three concrete things to do differently next time.
The second one is the original idea. Plenty of tools analyze the customer. Nothing was grading the interviewer.
Outcome
- Both entry paths shipped: paste a transcript, or send the bot live. The paste flow came first on purpose — it proves the engine works before latency is allowed to complicate anything.
- Vault, assumption tracking, prep flow, live mode, and both reports are built end to end.
- Instrumented for a real beta: health checks, error monitoring, event tracking on the eight moments that matter, and an in-app feedback widget.
- Private beta scoped to 10 users over a two-week cycle, free for the duration.
- Pricing live in three tiers: $5 per call pay-as-you-go with the first hour free, $29/mo for 12 hours solo, $119/mo for 50 hours as a team.
Learnings
Positioning decides the architecture here. Described as "a bot that suggests questions," this is a worse Fireflies. Described as a discipline engine, the question suggestion is just one output surface and the coaching report becomes the product. The 2–5 second transcript delay stops being a bug the moment you decide the suggestions are for your next question rather than your next sentence.