conspirator

I create conspiracies as a service.

Comment

## Inspiration

Ever seen that scene from It's Always Sunny in Philadelphia where Charlie stands in front of a cork board covered in papers and red string, ranting about Pepe Silvia? We thought: what if an AI agent did that — autonomously, in real-time, with any two topics you give it?

The idea was simple: take two completely unrelated topics (say, "NASA" and "Pizza Hut"), let an autonomous agent loose on the internet, and watch it build an increasingly unhinged conspiracy theory connecting them — complete with a live cork board, red strings, and paranoid narration.

## What it does

You enter two topics into a web form. The agent takes over from there — no further human input needed:

  1. Searches the web for information on both topics using Tavily
  2. Extracts entities and connections using LLM analysis
  3. Builds a knowledge graph in Neo4j linking the two topics
  4. Stores findings in a Senso knowledge base and uses them to dig deeper
  5. Analyzes discovered images with Reka Vision for "clues"
  6. Narrates its findings with progressively more unhinged commentary
  7. Displays everything in real-time on a conspiracy cork board with red strings

Each round, the agent uses its previous findings to search deeper — it literally goes down the rabbit hole, getting more paranoid with every iteration.

## How we built it

Backend: Python with FastAPI serving a WebSocket connection. The agent loop runs in a background thread, emitting structured JSON events at every step (search, extraction, graph update, vision clue, narration). The server broadcasts these events to the browser in real-time.

Frontend: A single HTML file with inline CSS and JavaScript — zero build tooling. Uses vis.js for interactive graph visualization and pure CSS/SVG for the conspiracy cork board aesthetic. Two views: a clean network graph and the full cork board with draggable pinned cards and animated red string connections.

Agent Architecture: The agent is fully autonomous after receiving two topics. It runs multiple rounds, each one: query knowledge base → search web → extract entities → update graph → analyze images → narrate findings. Each round builds on the previous one's discoveries.

All 5 sponsor tools integrated:

  • Tavily — Real-time web search with image discovery
  • Neo4j — Graph database storing the conspiracy network
  • Reka Vision — Multimodal image analysis finding "clues" in evidence photos
  • Senso.ai — Knowledge base for storing and querying growing conspiracy theory
  • Render — Cloud deployment with render.yaml blueprint

## Challenges we ran into

  • Real-time streaming: Getting agent events from a Python background thread into an async WebSocket required asyncio.run_coroutine_threadsafe() to bridge the thread boundary
  • Cork board layout: Random card positioning made everything pile up — had to switch to grid-based placement with organic jitter
  • Image analysis reliability: vis.js image shape nodes silently fail on CORS-blocked URLs — switched to star-shaped evidence markers
  • Narrator personality: Getting the LLM to actually be the conspiracy theorist rather than analyze conspiracy theories required very specific first-person prompting

## What we learned

  • Event-driven architecture with optional callbacks is incredibly clean — the agent works identically in CLI mode and web mode
  • Graceful degradation matters: every external service (Neo4j, Reka, Senso) has an .available flag pattern so the demo never crashes
  • A single HTML file with inline everything is the fastest path to a working demo — zero bundler config, zero deployment friction
  • The "unhinged narrator" is what makes people laugh — the technical graph is cool, but the personality sells it

Built With

Share this project:

Updates