-
-
Intro
-
How it works
-
Agents
-
Result
-
Vertex Ai Agents
Inspiration
The interview process is one of the most high-stakes moments in a person's career, yet most people practice alone in front of a mirror with no real feedback. AceIt was born from the idea that everyone deserves a personal interview coach, not just those who can afford one.
What it does
AceIt is a real-time AI interview coaching platform that conducts a full mock interview session with the user. Four specialized Gemini 2.5 Flash agents work simultaneously: the Speech Agent analyzes clarity and pacing, the Vision Agent reads body language and eye contact through the camera, the Coach Agent delivers spoken feedback after each answer, and the Voice Agent handles natural conversation flow. The system generates follow-up questions based on the user's actual answers, just like a real interviewer. At the end of every session, a full performance report is generated with an overall score, speech metrics, vision analysis, key strengths, and personalized areas for improvement.
How we built it
AceIt was built entirely inside Firebase Studio using Next.js and TypeScript for the frontend. Firebase Studio's AI assistant helped scaffold, debug, and iterate on the session page, results page, and AI flow integrations rapidly. The four AI agents were built using Vertex AI Agent Builder powered by Google Gemini 2.5 Flash. The backend is a Python FastAPI server handling the scoring and coaching endpoints. The app is deployed on free-tier cloud hosting with no infrastructure costs.
Challenges we ran into
Getting four agents to coordinate in real time without creating feedback loops was the biggest challenge. The microphone was picking up the AI's own voice output and transcribing it as user speech. This required building a synthesizing lock using a ref that pauses speech recognition exactly while the AI is speaking. Camera rendering in Next.js also required a specific fix where the media stream had to be stored in React state and attached to the video element through a separate effect to guarantee the ref was ready. Voice consistency was another issue as browsers load voices asynchronously, causing the first question to sometimes use a different voice than the rest of the session.
Accomplishments that we're proud of
Building a fully working real-time interview coach with live speech transcription, camera-based body language analysis, and spoken AI feedback all running together in a single session is something the team is genuinely proud of. The dynamic follow-up question generation makes the experience feel like a real conversation rather than a static quiz. The results page delivering a scored performance report with strengths and improvement areas gives users something concrete and actionable after every session.
What we learned
Real-time AI applications require much more careful state management than standard web apps. Race conditions between async AI calls, speech synthesis, and microphone input all need explicit coordination. Firebase Studio proved to be a powerful environment for rapid AI-assisted development, dramatically speeding up iteration. Vertex AI Agent Builder made it possible to deploy specialized agents quickly without managing complex infrastructure.
What's next for AceIt
The next step is expanding the question library to cover industry-specific interviews such as software engineering, product management, and healthcare. The team plans to add a progress tracking dashboard so users can see improvement across multiple sessions over time. Integration with calendar tools to schedule mock interview sessions and a mobile app version are also on the roadmap.
Log in or sign up for Devpost to join the conversation.