Master of Puppets

AI Hackathon Submission

4 AI products.
1 architecture.
0 hallucinations.

Four products people already use at work — request management, the directory, move planning, and in-app help — rebuilt so you can just say what you want. One principle underneath: the model proposes, our code decides.

Start walkthroughTeam

THE THESIS

The model proposes. Our code decides.

Every product in this submission obeys one rule: the model never mutates data, never invents identifiers, and never crosses a permission boundary. It chooses validated tools — our code executes them. This is the discipline that lets these ship to production, not just demo.

Loop depth 01Capability index 24%
Tool-callingThe model picks tools and arguments. Our tools resolve real IDs against the database and run the action.
Validated write portEvery write passes through a closed vocabulary. A hallucinated ID or unknown operation cannot get through.

How it holds up

Discipline you can count.

4products · 1 write architecture
100%of writes pass through a closed vocabulary
0model-invented IDs reach the database
2models, 2 jobs — Sonnet for tool-calling, Haiku for latency

The team

Built by Master of Puppets Team.

Luisdien
Fabián
Enrique
Diego

Four products. One architecture. Every write validated. Every permission scoped. Every count, name, and ID resolved against a real query — never invented by the model. This is what we mean when we say AI ships to production.

Claude is the reasoning brain (sonnet-4-6 for heavy tool-calling, haiku-4-5 for the latency-sensitive Find Bar). OpenAI handles voice — Whisper for transcription, gpt-4o-mini-tts (voice “nova”) for replies. The architecture underneath is ours.