Markfit

Inspiration

I went into Transform wanting to build something I'd actually be proud of, not just another idea chasing a prize. I kept circling ideas that sounded smart but had no real conviction behind them until I landed on something I actually live: founders can't afford real customer discovery. Talking to real people costs time and money most early-stage founders don't have, and asking a single AI chatbot "is this a good idea" just gets you a biased guess dressed up as feedback. That gap is real, and I wanted to build the honest version of solving it.

What it does

Two isolated AI agents have a real, unscripted conversation about your product. One agent pitches, grounded strictly in the product description you give it, nothing invented. The other plays a customer persona built from real objection data, not generic skepticism. They don't share context. Neither knows it's talking to an AI. You can tune the marketer's approach and the customer's relationship style on a live 2D matrix, and the output isn't just a transcript, it's a scored breakdown: objections, what worked, what didn't, and what to actually change.

How I built it

I started by proving the core mechanic alone in a terminal script before touching any UI, two separate message histories, two separate API calls, no shared context. Once that worked, I layered in real grounding: a persona system built on actual marketing methodology, not invented backstories, plus Reddit's public API pulling real language patterns into the mix. I locked the design system early, pure black, frosted glass, one accent color, and built every screen around that so the whole thing feels like one product instead of a patchwork.

Challenges I ran into

The hardest problem wasn't technical, it was getting the conversations to stop sounding like AI being agreeable with itself. Early versions resolved positive every single time, which made the whole tool useless as a validation signal. I had to explicitly force outcome diversity and ban AI-tell phrasing at the prompt level, fix isolation bugs where both agents were reading from one shared string instead of two, and rebuild the four "personality" corners of the matrix with actual distinct behavior instead of a label swap on the same generic voice. That fight for realism ate more of my time than any feature did, and honestly made the product better for it.

What's next

The RAG grounding is already wired, it just needs more real sources feeding it. Reddit's live, but Google Reviews and X/Grok integration would let the customer agent pull from a much wider, more current slice of how real people actually talk and complain, not just one subreddit. That's the version of this I actually want to ship.

Built With

• claude
• vercel