Prediction Market Forensic Monitor

A multi-workflow n8n system that polls Polymarket trades every five minutes, flags anomalous activity, runs AI forensic analysis, alerts via Discord, and logs everything to Supabase — with a weekly meta-agent that learns from its own misses.

The three workflows

Forensic Monitor & Strategy Agent

• Scheduled polling of Polymarket CLOB trades with adaptive lookback (6-minute delta vs. 8-hour backfill after downtime)
• Anomaly filter logic — niche snipers, macro whales, high-value sports trades, wallet clustering
• LLM-powered forensic analysis with risk scores and betting recommendations
• Discord alerts with market links and structured recommendations

Resolution Checker (Evaluation)

• Fetches unresolved predictions from Supabase forensic_logs
• Checks actual market outcomes via Polymarket Gamma API
• Marks predictions as resolved and records accuracy for the learning loop

Meta-Agent (Learning Engine)

• Weekly run that pulls wrong predictions from Supabase
• Formats failure context and feeds it back to the LLM to refine detection heuristics
• Closes the loop: monitor → predict → evaluate → learn

Infrastructure

• Docker Compose for local n8n + data persistence
• Supabase as the forensic log store and evaluation backbone

Why it matters

A full sense → decide → act → evaluate → improve agent loop built entirely in n8n — the kind of architecture I bring to production AI systems, just applied to prediction markets as a research sandbox.