Audeza

Audeza — an AI pitch coach that generates scripts for you deck, then watches and listens to you rehearse in real-time, coaching your delivery like a mentor in the room.

Comment

What Inspired me to build Audeza?

I've always been someone full of ideas. Over the years, I've built several of them — shipping products, writing code, solving real problems. But every single time, I hit the same wall: the pitch. I could build it, but I couldn't sell it. I'd stumble over my words, rush through slides, freeze when investors pushed back with hard questions. I knew my ideas were good — I just couldn't make other people feel that too. Mentors helped when they were available, but I couldn't call someone at midnight the day before a demo day. I needed a practice partner that was always there, that could actually watch me present and tell me what I was doing wrong. So I built one.

Audeza is an AI pitch coach that sees you, hears you, and coaches you in real-time. You upload your deck, turn on your camera, and start presenting — and Audeza gives you live feedback on your delivery, pacing, body language, and slide content. It can also flip into investor simulator mode, roleplaying as a tough VC who throws curveballs so you're never caught off guard again. After each session, you get a scorecard with per-slide breakdowns and pacing charts — so you can actually track yourself getting better.

Audeza was built with:

  • Gemini Live API for real-time multimodal streaming — audio, webcam, and slide awareness all at once
  • TanStack Start with file-based routing and server functions
  • Firebase Auth, Firestore, and Cloud Storage for the backend
  • Backend deployed on Cloud Run
  • React + shadcn/ui + Tailwind for a clean presenter-focused UI
  • Bun runtime with a WebSocket proxy bridging the browser to Gemini's live API
  • Audio via AudioWorklet at 16kHz PCM, webcam at 1 FPS JPEG — all streamed in real-time

Challenges I ran into:

  • Managing Gemini's context window with continuous audio and video — I had to design smart compression to work within the 2-minute media limit
  • Making slide awareness feel seamless so feedback is about this slide, not generic advice
  • Building a teleprompter that auto-scrolls by tracking your actual speech — matching spoken words to scripted text in real-time without being distracting.

Accomplishments that I'm proud of:

  • The coaching is truly multimodal — it watches your body language, hears your tone, and reads your slides all at once, just like a real mentor sitting across the table
  • Investor simulator mode that actually makes you sweat — the kind of practice that makes the real meeting feel easy
  • A scorecard that gives you honest, specific, per-slide feedback — not "great job!" but "you spent 45 seconds on slide 3 and said 'um' seven times"

What I learned ?

  • The Gemini Live API is incredibly powerful but demands careful orchestration when combining multiple media streams
  • Context-aware feedback (knowing which slide you're on, what it says) is what separates useful coaching from generic AI commentary
  • The hardest design problem wasn't technical — it was making real-time feedback feel helpful instead of interrupting your flow

What's next for Audeza ?

  • Launch publicly and get it into the hands of founders preparing for realdemo days
  • Team practice mode — let co-founders rehearse together with AI moderation
  • Industry-specific coaching — different feedback for VC pitches, sales demos, grant panels
  • Generate the slide decks
  • The bigger vision: no one should lose an opportunity because they couldn't communicate their idea.

The product was always good — the pitch just needed practice.

Built With

Share this project:

Updates