Skip to content

Syst.md

Syst.md is a voice-driven tool that turns a spoken system design conversation into both a structured Markdown spec and a live 3D architecture diagram.

Code Video Cerebral Valley

Core trick

A Gemini Live auditor red-teams each design section by voice, and a Zod-validated agent builds a Three.js scene from the output.

Tags

dev toolsvoice

Description

Syst.md is a voice-driven tool that turns a spoken system design conversation into both a structured Markdown spec and a live 3D architecture diagram. You talk to a conversational AI auditor (Gemini Live, real-time audio) that walks you through four sections in order: goals, architecture, components, and data flow. After each section is confirmed, the auditor red-teams it with one sharp critique, a failure mode, scaling cliff, or security gap, forcing you to harden the design before moving on. As sections are confirmed, DESIGN.md is written incrementally and a separate interpreter agent (Gemini structured output, Zod-validated) generates a spec that drives a Three.js scene with per-type shapes, ELK-laid-out edges, and particle animations showing data flow direction. The 3D diagram updates progressively after every section so you watch your architecture build itself in real time, with a 2D/3D view toggle and fullscreen. This solves three problems at once: design conversations are usually unrecorded, adversarial thinking is rarely applied during early design, and architecture diagrams are typically static drawings made hours after the discussion all three happen automatically as you talk. The final output is a hardened DESIGN.md plus an interactive 3D visualization of the system you described, produced in a single conversational session.
sskarz/Google-IO-Hackathon-2026★ 00 forksPythonCreated May 23, 2026 · last push May 24, 2026

README excerpt

# Google I/O Hackathon 2026

Voice-driven system design tool. Speak a description aloud; the app produces both a structured `DESIGN.md` and a 3D visualization of the system topology.

---

## Pipelines & Agent Capabilities

The platform operates on two distinct interactive pipelines:

### 1. The Real-time Interactive Gemini Live Auditor (`/auditor/ws`)
This bidirectional voice loop leverages the **Gemini Live API** (`gemini-3.1-flash-live-preview`) to conduct a real-time system design interview. 
- **Voice Loop**: Processes low-latency raw PCM audio packets streaming over a WebSocket connection.
- **Dynamic Hardening**: Features a 5-phase red-teaming cycle (Discuss, Confirm, Red-Team, Harden, Transition) across four sections: Goals, Architecture, Components, and Data Flow.
- **Tool-Calling Integration**: Live-calls `confirm_section` or `finalize_design` to rewrite the design and update the diagram in real time as the conversation progresses.

### 2. The Static Browser Voice Pipeline (`/generate-design` & `/generate-spec`)
A simpler, one-shot voice-to-design tool:
```
mic → Web Speech API (STT) → POST /generate-design (google-antigravity)
                                         │
                                         ▼
                                     DESIGN.md  ──────────────┐
                                         │                     │
                                         ▼                     ▼
                                    GET /design-md       POST /g

Team