Pet Committee

Three AI pets. One dilemma. Biscuit panics. Waffle says something accidentally wise. Marcus says the thing you've been avoiding. The committee is always in session.

Comment 1

Inspiration

After a long day, I usually find comfort in animals. I wanted to build something that felt warm and personal rather than just useful — so I created a committee of three AI pets who always have opinions about your problems.

What it does

Pet Committee is a voice-first multi-agent AI app. You type or speak your dilemma, and three pets respond with distinct personalities and voices:

  • Biscuit the chihuahua panics about worst-case scenarios
  • Waffle the golden retriever says something warm and accidentally wise
  • Marcus Aurelius the cat delivers the one observation you were avoiding

Each pet is a separate AI agent with its own system prompt, model, and ElevenLabs voice. There's also Chaos Mode — all three respond simultaneously. It is, by design, the worst possible way to receive advice.

How I built it

  • Groq (Llama 3.3 70B) powers Biscuit and Marcus
  • Google Gemini 2.5 Flash powers Waffle — chosen specifically because Gemini's conversational warmth matched his accidentally-wise personality better than other models
  • ElevenLabs TTS gives each pet a distinct voice (Jessica's anxious quirk for Biscuit, Bill's warm resonance for Waffle, George's dry detachment for Marcus)
  • Next.js 15 + Tailwind for the frontend
  • Aseprite for all pixel art — the room background and three animated pet spritesheets
  • Deployed on Vercel

Challenges

  • Getting three AI agents to sound genuinely different required 5+ rounds of prompt iteration per character — especially Waffle, who kept drifting into therapy-speak
  • Vercel's serverless filesystem is read-only, which broke the ElevenLabs audio caching approach that worked locally — had to switch to direct streaming
  • Balancing the pixel art style with the warm cozy aesthetic while keeping animation frames manageable within the hackathon timeline

What I learned

  • Multi-agent systems need strong character separation at the prompt level — "never say" lists and explicit rhetorical constraints matter as much as the positive instructions
  • Voice is not decoration in a voice-first app — it carries emotion that text alone can't
  • Gemini and Groq have genuinely different "personalities" that suit different characters

Built With

  • aseprite
  • elevenlabs-tts
  • google-gemini-2.5-flash
  • groq
  • llama-3.3-70b
  • next.js
  • tailwind-css
  • typescript
Share this project:

Updates