Building Candor: AI-Powered Product Discovery at the Gemini Live Agent Challenge
The Problem I Wanted to Solve
I’m a builder, and the hardest part of building good software isn’t the code; it’s listening. Specifically, listening to what users are actually struggling with, not just what they say they want.
AI already helps me write code faster. I wanted it to help me with the part that matters more: understanding the problem before I build the wrong solution.
What Is Candor?
Candor is an AI-powered digital twin that lets product managers and developers run product discovery interviews without needing a real customer in the room.
The research backs this up: LLMs trained on demographic personas reproduce human consumer responses with 90% test-retest reliability, making them effective stand-ins for early-stage customer discovery.
Here’s how it works:
- You describe a customer persona
- You conduct a voice interview, and you can interrupt the AI mid-response, just like a real conversation
- The AI surfaces pains, gains, and jobs to be done as a real customer would
- Hesitations, workarounds, and emotional cues are captured as signals
- Follow-up questions are suggested to help you dig deeper
How I Built It
The technical architecture was one of the most interesting parts of the project:
- Event-sourced domain model -every insight, association, and session event is immutable and replayable
- Dual projections -events project into both an RDF knowledge graph (SPARQL-queryable) and relational tables (SQL-queryable)
- Async enrichment pipeline -a dedicated Gemini Flash model extracts structured insights after each conversation turn without blocking the flow
- Embedding-based deduplication -cosine similarity at a 0.85 threshold catches paraphrased duplicates so insights don’t pile up redundantly
- Dynamic voice selection -an LLM picks from 8 voice profiles to match the persona description
- Deployed on Google Cloud Run, with prompts fine-tuned using promptfoo
The stack: Python, Gemini, and the Agent Development Kit (ADK).
The Pivot
Initially it was built from the perspective of a user using the tool to describe an issue they were having and the AI would help the user understand the problem better. But I quickly realized that the real value was in using the tool to run discovery interviews with other people. So I pivoted to make it a tool for running interviews, and that turned out to be a much more compelling use case.
What I’m Proud Of
I shipped a working product in 4 days -and then actually used it to validate other ideas I’d been kicking around. That’s the best proof of concept: when the tool you built becomes the tool you reach for.
The live voice experience was also a revelation. You take for granted the subtleties of voice-enabled apps -like allowing the user to cut off the AI mid-sentence -but building it from scratch made me appreciate every nuance.
Another unexpected takeaway: this way of running discovery interviews is surprisingly forgiving. Without the social pressure of keeping a real conversation flowing, you have the space to think, probe, and iterate freely.
What’s Next
I’d love to integrate Candor with common product discovery tools and make it easy to drop into different workflows. If you work in product and want to try it, the app is live: