Technical Architecture
A complete walkthrough of the autonomous 6-stage research pipeline, from raw intelligence collection to final portfolio consensus.
This platform is built on the TradingAgents framework from Tauric Research, an open-source agentic AI library that orchestrates large language models into specialized analyst roles using a directed graph architecture. Each AI agent operates autonomously within a constrained domain, contributing its independent analysis to a shared state object. This is not a chatbot. This is a structured, multi-agent reasoning system where AI models are assigned roles, given context, and forced to produce defensible analysis under adversarial conditions.
The platform has been significantly extended by Brian Galvan, an AI researcher and developer, with several proprietary systems that go well beyond the original framework:
OSINT Intelligence Layer (Stage 0) autonomously sweeps 12 public data channels, including SEC filings, federal contracts, Congressional trading disclosures, options flow, insider transactions, macroeconomic indicators, and retail sentiment. This intelligence is compiled and injected directly into the agent pipeline, ensuring every AI analyst, debate argument, and risk assessment is grounded in real-world data.
Crypto Intelligence Analyzer activates a specialized agent, Jake Summers, when the target is a digital asset. He executes on-chain analytics, whale wallet tracking, exchange flow monitoring, DeFi protocol TVL analysis, and crypto-native social sentiment aggregation to produce a crypto-specific intelligence briefing.
Polymarket Prediction Pipeline deploys 6 adversarial agents purpose-built for binary prediction markets: an intelligence analyst gathers event context, YES and NO advocates construct opposing probability arguments, a contrarian analyst stress-tests for cognitive bias, a probability synthesizer calibrates the final estimate, and an edge calculator compares the AI's probability against live market prices to identify mispriced contracts.
SimTrader is an autonomous simulation engine that executes paper trades using the full pipeline in real time against a $100K portfolio, tracking P&L to validate research accuracy. DayTrader compresses the pipeline for intraday speed, scanning momentum shifts, VWAP deviations, and order flow imbalances to generate actionable entry and exit signals within minutes.
The result is a fully autonomous agentic workflow spanning equities, crypto, and prediction markets, where AI agents independently research, argue, judge, and assess risk before a single recommendation is produced. The only human in the loop is at the final decision stage.
Before a single analyst agent is activated, the system executes a full-spectrum open-source intelligence sweep across multiple classified data verticals. The reconnaissance engine queries regulatory filing systems, federal procurement databases, macroeconomic data feeds, institutional disclosure networks, derivatives positioning data, and retail sentiment aggregators in parallel. The resulting intelligence briefing is compiled into a structured document that serves as the foundational input for every downstream agent in the pipeline.
Four independent AI analyst agents process the intelligence briefing through their specialized domains: technical analysis, sentiment intelligence, news/media processing, and fundamental valuation. Zero cross-communication ensures unbiased, independent assessment.
Two adversarial AI agents construct opposing investment theses from the same data. Neither is allowed to equivocate. This forces the system to surface risks and opportunities that consensus-seeking approaches miss entirely.
Asymmetric upside, catalysts, accumulation signals
Downside scenarios, valuation risks, competitive threats
An AI Research Director evaluates both sides without predetermined bias. The verdict determines which thesis is more defensible, establishing the intellectual framework for the final decision.
Three AI risk analysts examine every position through aggressive, neutral, and conservative lenses. A Risk Judge synthesizes all inputs into a unified risk verdict.
All AI research, debate outcomes, and risk assessments converge here. Every position change and risk parameter is validated against the complete analytical record. The only human in the loop.