Skip to content

anilandcode/autopsy

Repository files navigation

AUTOPSY

Forensic Intelligence — 4 Modes, 24 Agents, 1 Verdict

Autopsy

Live Demo: autopsy-nine.vercel.app

What it is

Autopsy investigates why companies fail — and what could have saved them. Six specialized AI agents per mode research in parallel, debate each other's findings, and produce a forensic verdict in ~22 seconds.

Four modes:

  • Postmortem — investigate why a company failed
  • Pre-Mortem — predict what could kill a living company
  • Founder Mode — analyze your own startup before it fails
  • Counterfactual — explore alternate histories: "What if they made a different decision?"

Counterfactual Mode

The 4th mode asks: What if [company] had made a different decision?

Six counterfactual agents reason about what didn't happen:

  • CF Market Analyst — skeptical about internal decisions changing external market realities
  • CF Operator — assesses whether the alternate decision could actually be executed
  • CF Money Trail — models the financial impact of the alternate path
  • CF Customer Voice — evaluates whether users would have responded differently
  • CF Engineer — honest about technical complexity vs. leadership beliefs
  • CF Historian — finds real precedents (requires 2+ historical cases as evidence)

Each agent must: understand the actual causal chain, identify the decision point, model the alternate chain, find real precedents, and assess second-order consequences (butterfly effects).

The synthesizer renders one of five verdicts: would have survived, would have delayed failure, would have failed differently, would have made no difference, or could have transformed the company.

Preset scenarios: Blockbuster/Netflix, Kodak/Digital, Yahoo/Google, Quibi/TV, Theranos/Real Science, MySpace/Better Tech.

Why AMD MI300X

192GB HBM3 lets us load all 6 agents (each ~70B parameters) simultaneously. On a single H100 (80GB), this requires 3 sequential rounds. On MI300X, it runs in one parallel pass — enabling real-time agent debate that wasn't possible before.

H100 MI300X
Memory 80 GB HBM3 192 GB HBM3
Agents 3 sequential rounds 1 parallel pass
Debate Impossible Real-time
Time ~75s ~22s

Architecture

See the full architecture breakdown →

Design System

  • Palette: Indigo (#4B4BA0) and violet (#8F47AE) accents on near-black (#0F1110) surfaces
  • Typography: Newsreader serif for display, Inter for UI, JetBrains Mono for data
  • Surface language: Bordered cards with border-white/5, gradient glow shells, and subtle corner brackets
  • Background: Three.js data-wave mesh with dual-layer crossfading video, overlaid with solid tint + bottom-to-top gradient for text legibility

Tech Stack

  • Next.js 15 + TypeScript + Tailwind v4
  • Kimi K2.6 via Fireworks AI (262K context window)
  • Tavily Search API for evidence gathering
  • Server-Sent Events for streaming agent updates
  • Vercel edge deployment

Local Development

# Clone
git clone https://github.com/anilandcode/autopsy.git
cd autopsy

# Install
npm install

# Environment
cp .env.example .env.local
# Add your API keys:
# FIREWORKS_API_KEY=     (for Kimi K2.6 via Fireworks AI)
# TAVILY_API_KEY=        (for web search)

# Run
npm run dev

Open http://localhost:3000 locally, or visit the live demo at autopsy-nine.vercel.app.

License

MIT

Built for

AMD Developer Hackathon 2026 by lablab.ai

About

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors