Inspiration
I recently made the switch from software development to product management and realized PM interviews are a completely different game. There's no binary pass/fail, it's conversational, ambiguous, and the feedback loop barely exists. The options are limited: practice with friends who sugarcoat, or rehearse alone with no feedback at all. Neither tells you where your answer actually broke down and whether you are improving or not. Primed was built to fill that gap.
What it does
Primed is an AI interview coach that conducts mock interviews, evaluates the user's performance, and gives them unbiased feedback. It runs real-time voice-based practice sessions, generates causal feedback that pinpoints the exact moment an answer fell apart, maintains an evolving skill profile across core competencies, and autonomously selects the next drill based on identified weaknesses.
How we built it
The system is built around two Gemini-powered agents.
The first is the Voice Interview Agent, built with Google's Agent Development Kit and Gemini Live API. It handles real-time, adaptive conversation
The second is an Orchestrator Agent that manages three sub-agents:
- Feedback Agent that analyzes the full transcript to generate causal, skill-level feedback
- Memory Agent that maintains an evolving user profile summary, capturing patterns in how the user succeeds and fails across sessions
- Recommendation Agent that selects the next practice session using the updated skill profile, user context, and available drill library
Challenges we ran into
The biggest challenge was making the voice agent proactive rather than passive. We didn't want it to just respond, it needed to initiate the conversation, drive the interview forward, and naturally wrap up the session when it had enough signal. We achieved this through tool calls that give the agent control over session flow, including ending the call on its own. Beyond that, prompt tuning was an ongoing effort, getting the agent to probe, push back, and stay within the scope of the drill without feeling robotic or overly agreeable.
Accomplishments that we're proud of
An early user used Primed to prepare for a PM interview and cleared the screening round. That's the validation that matters most for us, the system works in a real-world setting, not just as a demo
What we learned
PM interview feedback is a surprisingly hard problem. The difference between useful and useless feedback comes down to specificity, connecting observations to exact moments in the conversation. We also learned that the evolving user profile is what makes the system compound over time, it shifts from referencing specific past sessions to recognizing behavioral patterns
What's next for Primed
The long-term vision is to turn Primed into a holistic interview prep agent, not just mock interviews, but the entire pipeline. That includes discovering relevant job opportunities, fine-tuning resumes for specific roles, and running targeted mock interviews based on the company and position. We also plan to expand beyond product management into other interview domains like marketing, design, and engineering. The architecture supports this, the skill rubric and drill library are modular, and the evolving user profile already captures background, strengths, and patterns, which naturally extends to resume personalization and job matching.
Log in or sign up for Devpost to join the conversation.