PMF Agent

## Inspiration

This project started from a very specific and common startup pain: **customer discovery calls that feel productive in the moment but fail to produce clear direction afterward**.

Founders know they should be talking to customers, but in practice those calls are hard to run well. During a single discovery conversation, a founder is often:

- trying to deeply listen and build trust  
- mentally mapping what the customer is saying to product hypotheses  
- deciding what to ask next  
- resisting the urge to pitch their solution  
- taking notes for later  

Because of that cognitive load, discovery calls often drift. Founders ask **leading or shallow questions**, miss important follow-ups, or move too quickly into solution mode. Different team members also run interviews differently, which leads to **fragmented signals** and conflicting interpretations.

The result is a slow, noisy path to Product–Market Fit—and teams end up shipping features based more on intuition and vibes than on validated customer insight.

I was inspired to build something that could sit quietly in the background of a **customer discovery call**, helping founders stay in discovery mode and ask better questions at the moments that matter most.

---

## What I Learned

Building this project taught me that **the highest leverage moment in customer discovery is not analysis after the call—it’s the follow-up question during the call**.

A single well-timed prompt—asking about a workaround, frequency, impact, or willingness to pay—can radically change the quality of insight you get. But those are exactly the questions founders miss when they’re juggling execution pressure and conversation flow.

I also learned that customer discovery isn’t about collecting more data—it’s about **maintaining focus on hypotheses**. When real-time guidance is tied to what you’re trying to learn (problem severity, alternatives, urgency, pricing sensitivity), interviews become more consistent and comparable across a team.

Finally, I learned that the best tools for discovery don’t replace the human. They reduce mental overhead so the founder can listen better, stay curious longer, and avoid defaulting to pitching. When that happens, customer conversations stop being anecdotal and start becoming a reliable input into PMF decisions.

How We Built It

We built this project by combining lightweight frontend tooling with AI services that could work reliably in real time during customer discovery calls.

On the frontend, we used Claude Code and Cursor to rapidly prototype and iterate on the interview experience. These tools helped us move quickly from idea to working UI, especially when wiring up live audio capture, managing transcript state, and triggering question generation without breaking the flow of a call.

For the backend intelligence, we used Yutori to power the AI system responsible for generating follow-up questions. Yutori helped us create an agent that takes live conversation context and produces focused, non-leading questions aimed at uncovering real PMF signals such as problem severity, current workarounds, and urgency. This allowed us to keep the “thinking” layer flexible and grounded in discovery goals rather than generic chat responses.

To handle audio transcription, we integrated ElevenLabs, which converts live audio from customer discovery calls into text that can be processed downstream. Accurate transcription was critical, since question quality depends heavily on capturing the customer’s exact language and intent.

For reporting and synthesis, we used the Freepik API to generate clean, visual reports once interviews were completed. These reports help teams quickly scan findings, themes, and patterns across interviews, making discovery insights easier to share with the rest of the startup.

Because real customer data is sensitive and limited early on, we used Tonic.ai to generate fabricated but realistic interview data. This allowed us to safely test reporting flows, summaries, and insight visualizations without exposing real customer conversations, and helped us validate the end-to-end experience before deploying with real users.

Overall, the system was built as a modular pipeline—frontend capture, transcription, AI-driven question generation, and post-call reporting—so each piece could evolve independently while supporting the core goal: helping startups run better customer discovery calls and reach PMF faster.

Challenges We Faced

One of the biggest challenges was refining the idea itself.

Customer discovery is a broad space, and early on it was tempting to build a tool that did everything: transcription, analytics, dashboards, summaries, insights. The hard part was narrowing the focus to the highest-leverage moment—the live customer discovery call—where founders most often lose valuable insight. Defining the problem clearly as “help founders ask better follow-up questions in real time” required multiple iterations and conscious restraint.

Another major challenge was defining the problem precisely enough to build against it. Startups don’t fail because they lack data; they fail because conversations are inconsistent, biased, and fragmented. Translating that abstract problem into concrete product behavior—what should the tool do during a call versus after it—took careful thought and experimentation.

On the technical side, integrating the frontend with the rest of the system was non-trivial. Real-time audio capture, live transcription, backend storage, rolling summaries, and question generation all had to work together without introducing latency or disrupting the flow of a live conversation. Small mismatches between frontend state and backend context could easily lead to irrelevant or delayed prompts, which defeats the purpose of real-time guidance.

Balancing responsiveness, context accuracy, and a clean user experience was critical. The tool had to feel lightweight and supportive, not distracting. Solving that coordination problem—across frontend, backend, and multiple AI services—was one of the most meaningful challenges of the project.

Built With

• claude
• cline
• css
• cursor
• fastapi
• python
• tonic
• typescript
• yutori