CryptoFintech Digital Clone
Back to Projects

DePredict

Active

A multi-expert AI consultation framework that assembles domain-specialized AI agents for structured deliberation, helping users calibrate probability forecasts through organized debate on prediction markets.

Yingzhe Yu, Kani Chen
predictiondigital cloneNLP
WebsiteGitHub

How It Works

1

Knowledge Retrieval

Gather domain knowledge via RAG with YouTube transcripts and Tavily news, then partition information across experts to ensure independent perspectives.

2

Expert Assembly

Assemble 5-10 domain-specialized AI agents with varying analytical stances for structured cognitive conflict.

3

Structured Debate

Run a 3-round debate pipeline: independent analysis, cross-rebuttal, and final synthesis among all expert agents.

4

Aggregated Forecast

Aggregate predictions using 9 mechanisms including LMSR markets, reputation-weighted scoring, and Bayesian Truth Serum.

Live Preview

https://depredict.net/

Loading preview...

Visit Full Site

Tech Stack

PythonFlaskVue.jsDeepSeek V3RAGDocker

Overview

DePredict tackles a fundamental limitation of single-LLM forecasting: the lack of cognitive diversity. Instead of asking one model for a prediction, DePredict assembles a panel of domain-specialized AI agents that engage in structured deliberation — mimicking how expert panels operate in the real world.

The system creates "structured cognitive conflict" by partitioning available information across experts, ensuring each agent develops an independent perspective before engaging in cross-examination. This design prevents groupthink and surfaces a wider range of analytical angles.

Key Features

  • ✓ Assembles 5-10 domain-specialized AI agents with distinct expertise and analytical stances
  • ✓ 3-round debate pipeline: independent analysis, cross-rebuttal, and final synthesis
  • ✓ Interactive consultation mode for free-form dialogue with individual experts
  • ✓ 9 aggregation mechanisms including LMSR prediction markets, reputation-weighted scoring, and Bayesian Truth Serum
  • ✓ Information isolation (40% shared / 60% private) prevents groupthink
  • ✓ Supports cryptocurrency and sports prediction markets with extensible architecture

    • Performance

  • ✓ Model Edge: +0.0568 over baseline
  • ✓ Win Rate: 72.2% across evaluated markets