AUTOPSY: AI Business Failure Prediction System

Every business that failed had the warning signs. AUTOPSY reads them 6 months before anyone else does.

Comment

Inspiration

Every year, millions of businesses collapse. Not because the founders were incompetent, but because the warning signs were invisible until it was too late.

I kept asking one question: What if a business could see its own death coming?

Blockbuster had the data. Toys R Us had the signals. WeWork had the red flags. Nobody connected the dots in time.

That obsession became AUTOPSY.

What it does

AUTOPSY is an AI-powered Early Warning System for Business Failure. It monitors a company's public and internal signals in real time and generates a continuously updated Survival Score (0 to 100) with a predicted failure timeline and a prescriptive recovery plan.

It does not tell you how your business is doing today. It tells you when your business is going to collapse, and exactly how to prevent it.

How we built it

Survival Score Engine

  • Ingests financial data, customer churn signals, market trends, and competitor activity
  • Produces a weekly Survival Score with a probability curve
  • Example output: "At current trajectory, 73% chance of critical distress within 5 months"

Multi-Agent Signal Detector An intelligent agent network that continuously analyzes:

  • Cash flow burn rate anomalies
  • Customer review velocity drops
  • Competitor pricing movements
  • Google Trends shifts for core product categories
  • LinkedIn headcount changes (public early attrition signal)
  • News and social sentiment analysis

Prescription Engine AUTOPSY does not just diagnose. It prescribes:

  • A ranked, time-sensitive action plan
  • Example: "These 3 actions taken in the next 30 days improve your Survival Score by 18 points"

War Room Dashboard A cinematic real-time command center with:

  • Red / Amber / Green threat zone indicators
  • Failure probability timeline charts
  • Signal breakdown by category
  • AI chat interface: ask "Why is my score dropping?" and get a real answer ## How I Built It

Architecture Overview

$$\text{Survival Score} = f(\text{Financial Health}, \text{Market Signals}, \text{Sentiment}, \text{Operational Signals})$$

The score is a weighted composite model combining:

\( S = w_1 \cdot F + w_2 \cdot M + w_3 \cdot C + w_4 \cdot O \)

Where:

  • \( F \) = Financial Health Index (burn rate, revenue trend)
  • \( M \) = Market Signal Index (Google Trends, competitor activity)
  • \( C \) = Customer Sentiment Index (review velocity, NPS trend)
  • \( O \) = Operational Signal Index (headcount changes, news sentiment)

Stack Breakdown

Layer Technology
Frontend Next.js, Tailwind CSS, Framer Motion
Backend FastAPI, PostgreSQL
AI Agents LangChain multi-agent framework, GPT-4o
Forecasting Facebook Prophet, scikit-learn survival analysis
Data Sources SerpAPI, NewsAPI, LinkedIn public data, user CSV upload
Vector Store ChromaDB for RAG document context
Auth and Storage Supabase

Agent Architecture

AUTOPSY runs 4 specialized AI agents in parallel:

  1. Financial Agent - Processes uploaded financial data and computes burn indicators
  2. Market Agent - Pulls and interprets Google Trends and competitor signals
  3. Sentiment Agent - Analyzes news and review data for narrative shifts
  4. Synthesis Agent - Aggregates all agent outputs into the final Survival Score and Prescription Plan

Challenges we ran into

  • Data heterogeneity: Combining structured financial CSVs with unstructured news and social data required careful normalization pipelines
  • Signal lag: Public signals like LinkedIn headcount changes lag real events by weeks, requiring the model to account for temporal offsets
  • Explainability vs accuracy tradeoff: Making the AI explainable to non-technical business owners without sacrificing model depth
  • Solo development: Building a full-stack AI system with a multi-agent backend, cinematic frontend, and predictive engine solo within the hackathon window required ruthless prioritization

Accomplishments that we're proud of

Here is your Accomplishments section, clean and ready to paste:

markdown## Accomplishments That We're Proud Of

Building a fully functional multi-agent AI system as a solo developer within a hackathon timeframe is something we did not take for granted. Each agent operates independently, processes a different category of signals, and feeds into a unified Survival Score in real time. Getting that pipeline to work reliably end-to-end was a milestone in itself.

We are proud that AUTOPSY does not just predict, it explains. Most predictive systems output a number and leave the user confused. AUTOPSY tells you the score, shows you exactly which signals drove it down, and hands you a ranked prescription to reverse the damage. That explainability layer took the most iteration to get right, and it is the part that makes the product genuinely usable by a non-technical business owner.

Validating the model against historical business failures was a breakthrough moment. Watching AUTOPSY score Blockbuster at 21 out of 100 in 2008, correctly identifying Netflix-driven Google Trends collapse and abnormal LinkedIn attrition as the top two signals, using only data that was publicly available at the time, gave us confidence that the underlying approach is sound.

We are also proud of the War Room Dashboard. We wanted it to feel urgent, not decorative. The decision to borrow the visual language of a medical emergency room crossed with a military operations center was deliberate. When a user sees their Survival Score drop into the red zone, it should feel like an alarm going off, because it is.

Finally, building something that did not exist before. Business intelligence tools are everywhere. Reactive dashboards are everywhere. A system that continuously autopsies a living business and tells it how close it is to flatlining, with a path back, did not exist in this form. That is what we set out to build, and that is what AUTOPSY is.

What we learned

  • Building multi-agent systems that run in parallel and synthesize results is genuinely hard but the output quality is incomparable to single-model approaches
  • Survival analysis from medical statistics (Kaplan-Meier estimators) translates surprisingly well to business failure modeling
  • The most powerful part of the system is not the score itself, it is the explanation layer that tells users why the score is what it is
  • Historical business failures (Blockbuster, Toys R Us, WeWork) are goldmines for validating predictive models because the ground truth is known

What's next for AUTOPSY: AI Business Failure Prediction System

  • Integration with live accounting tools (QuickBooks, Xero)
  • Industry-specific survival models (retail vs SaaS vs manufacturing)
  • Investor-facing mode: due diligence reports generated automatically
  • Mobile alert system for real-time Survival Score drops
  • Expanding the historical validation dataset to 500+ documented business failures

Built With

Share this project:

Updates

Submission history