Inspiration

At every hackathon, someone asks: "Did you validate it?" Almost nobody can answer honestly. Customers are polite. They say "cool idea" and "I'd use that" even when they'd never pay. First-time founders hear "yes" and build for months. 43% of VC-backed startups that shut down since 2023 cited poor product-market fit (CB Insights, 2026). We wanted something that tells you, right after the call, what was real.

What it does

You paste a customer interview. ValiDate reads every sentence the customer said and marks it as real interest or just being polite. Then it gives a verdict: keep going, narrow down, try a new angle, or pivot. Every verdict comes with the customer's own words as proof. It also tells you what you did wrong (talking too much, pitching too early, leading questions) and the 3 questions to ask next. If there isn't enough evidence, it says so.

You can also talk to it on ASI:One: @validate.

How we built it

Our rule: Jev judges, OpenAI writes, Python counts.

Jev (TypeSafe AI) judges every customer sentence: commitment, past pain, hypothetical, compliment or small talk.
Python turns that into a score from 0 to 100:

$$\text{score} = 100 \cdot \frac{R}{R + P}$$

where $R$ adds up the real signals and $P$ the polite ones, each weighted by type and by how sure Jev is.

OpenAI writes the read-out in plain English, only from Jev's output. It never changes the verdict.
Next.js + Postgres for the app, FastAPI for the analyzer, all on Railway.
Four Fetch.ai agents on Agentverse: Echo listens, Mark marks every sentence, Tiki plans your next move, Val gives the verdict.

Challenges we ran into

Enthusiasm isn't evidence. "I'd 100% pay for that" sounds great and means little. A flat "send me the link, I'll try it Monday" means a lot. Getting that right was the hard part.
AI likes to paraphrase. We added a check that drops any quote that isn't word for word in the transcript.
Knowing when to say "I don't know." We built a gate so short or one-sided interviews get "not enough evidence yet" instead of a guess.

Accomplishments that we're proud of

A polite interview and a real-pain interview of the same idea get clearly different verdicts.
Every verdict shows its receipts.
It works end to end, on the web and inside ASI:One.

What we learned

Ask about the past, not the future. What people did beats what they say they'll do.
Politeness isn't the signal. A real commitment can sound polite, and a blunt line can be empty.
One model to judge, another to write. Mixing the two makes both worse.

What's next for ValiDate.ai

Audio and video upload, then recording Zoom and Meet calls directly.
A report across all your interviews, so you can see the evidence move.
The Validation Hour: one hour before submission at hackathons, every team interviews mentors and ships with evidence.
Cohort licenses for universities and accelerators.