PolyM.AI.rket

PolyM.AI/rket turns any binary question into a structured intelligence briefing : the kind that used to require a team of human analysts. Type a question. The system researches it live, then assembles a panel of 6–8 expert agents generated specifically for that question: a Fed chair for monetary policy, a RAND strategist for geopolitics, a VC and a regulator for tech governance. No fixed pool : the right experts for the right question, every time. The panel debates in two rounds. Round 1: each agent reasons independently from a shared fact sheet, outputs a probability and argument. Round 2: agents see only the consensus price — not each other's positions — forcing genuine conviction over social conformity. Aggregation uses logarithmic pooling (not a simple average), which corrects for overconfidence and produces a calibrated 80% confidence interval. On resolved historical questions, the ensemble scores a Brier score of 0.091 vs 0.148 for the best single agent — a 38% improvement. The final synthesis names who disagreed, why, and what new information would move the price. Built on AMD MI300X via AMD Developer Cloud. Qwen3-27B-FP8 on ROCm + vLLM serves 6–8 concurrent agents in parallel without memory pressure. Orchestrated with LangGraph, grounded with Tavily live search, streamed to a custom Slack-clone interface in real time.