Kuan Yu
← Journal

Gemini Hackathon Post-Mortem: Lessons for Next Time

Built a working voice AI demo in 7 hours. Here's what I learned about hackathon strategy.

TECH

Quick Context

Night before: Reviewed past winners with teammates. Planned prediction markets project. 12am discovered track themes exist. Night prep

Morning chaos: Car ride

  • Left JB at 7am, reached Google office 8am
  • Teammates rejected at entry (didn’t register individually)
  • Niels: “FOUNDERS MODE!” - used rejection as motivation Founders mode
  • Got in eventually. Drained before we started.

Setup failures:

  • No breakfast, no table, no seat
  • WiFi wouldn’t connect → couldn’t claim API credits
  • Lost first hour to logistics instead of building

End of day failure:

  • Left only 20 mins for video recording AND deployment
  • Didn’t know how to use my video tool
  • Tried to host to Cloud Run in the same 20 mins
  • Submitted with no demo video

What I Built

Meeting Prep Agent - Voice-triggered briefings before meetings

"Prepare me for my next meeting"
  ↓
Research attendee (Google Search)
  ↓
Generate bio card
  ↓
Voice briefing
  ↓
"Send confirmation email?"

Deployed to Cloud Run. Working demo. Only one who shipped.

Vision vs Execution (40% Shipped)

The idea was S-tier. The execution was B-tier. Here’s what I planned vs what I actually built:

What shipped:

  • ✅ Voice Output (TTS) - Google Cloud TTS works
  • ✅ Function Calling - Gemini orchestration works
  • ✅ Web Research - Serper API integration works

What didn’t:

  • ❌ Calendar Monitoring - Mock data, no real Google Calendar OAuth
  • ❌ Gmail History - Mock JSON, no OAuth
  • ❌ Bio Card (Nano Banana) - Didn’t implement image generation
  • ❌ Proactive Triggers - User-initiated only, not “before you ask”
  • ❌ Send Email Action - Demo shows it, doesn’t actually send
  • ⚠️ Voice Input - Web Speech API fallback, not Gemini Live

Execution rate: ~40%

The Bigger Vision I Had

Not just meeting prep - a mentor/coach that:

  • Tracks your goals across relationships over time
  • Remembers what you committed to in past meetings
  • Suggests strategic asks (“You’ve met Sarah 4x. Ask for the intro this time.”)
  • Proactively briefs you before you even open your calendar

This is fundamentally different from Gong/Fireflies. It’s a relationship operating system, not a meeting recorder.

The problem wasn’t the idea. It was 7 hours.

Product Instinct vs Hackathon Instinct

I kept envisioning a real user the entire time. How would someone use this? What would make them say “this is useful”? I optimized for immediate user benefit, not judge impression.

Different game, different rules. Good to know for next time.

Gap to S-Tier

Here’s why I ranked Tier B-:

  • Tools Used: 5 vs 25+ (Neuroflix)
  • Architecture: Single agent, vanilla JS vs Multi-agent orchestration
  • Demo: No video vs Full production video
  • Novel Tech: API wrapper vs EmergentDB’s custom algorithm in Rust
  • Engineering: No tests, flat files vs TypeScript, CI/CD, Vitest

What S-Tier Had That I Didn’t

  1. Breadth over depth - Winners showcased MORE Google tools, not deeper integration
  2. Multi-agent systems - Neuroflix had 8 specialized agents (Director, Scriptwriter, Editor…)
  3. Demo video - Judges watch videos, not live demos
  4. Novel algorithm - EmergentDB invented MAP-Elites for vector search. I assembled APIs.
  5. Enterprise architecture - EarthLinks had TypeScript + tests + CI/CD

The Takeaway

Hackathons and products optimize for different things.

Hackathon success factors:

  • Tool breadth over depth
  • Video polish matters as much as code
  • Demo storytelling is a core skill

Next Time Checklist

Before the Event (Race Day Rules)

Like a marathon - don’t wear new shoes on race day. All setup done BEFORE:

  • API credits claimed and tested
  • All tools familiar (no learning during competition)
  • Video recording tool practiced (I didn’t know how to use mine)
  • Hosting/deployment pipeline ready (don’t deploy on deadline)
  • Fallback data prepped in case APIs fail
  • Demo script outline drafted

Day Of

  • Write the 60-second demo script FIRST
  • Work backwards from demo to code
  • Use ONLY familiar tools (Claude Code > AI Studio)
  • Scope down until embarrassingly simple
  • Allocate last 2 hours for VIDEO, not features

What Wins

  • Maximize tool count (breadth > depth)
  • Multi-agent > single agent
  • Polished video > working code
  • Novel algorithm > API wrapper
  • Memorable demo moment > practical utility

Group photo

Resources