Origin - Black-Box Arena

Inspiration Financial markets are drowning in data, but starved for synthesis. Traditional tools like Bloomberg and Refinitiv are passive data silos—they provide the numbers, but a human must decide if they matter. Meanwhile, modern AI research assistants can summarize news, but they cannot synthesize a novel trading strategy, manage portfolio risk, or execute a trade. We asked ourselves: What if the AI wasn't just the analyst, but the entire investment committee? Inspired by the "Agentic Turning Point" of 2025, we built Origin - Black-Box Arena. We are moving the industry from "searching for data" to "harvesting synthetic alpha" by deploying autonomous agents that debate, allocate capital, and execute trades in milliseconds.

What it does Origin - Black-Box Arena is an AI-native hedge fund engine that converts real-time market events into structured, risk-aware trade decisions. Instead of acting as a passive research assistant, the system runs a live, multi-agent investment committee: The Event Trigger: Ingests real-time catalysts (earnings, macro data, news) via WebSocket streams. The Agent Council: Specialized Gemini agents (Bull, Bear, Risk Officer) debate the trade using real-time market data, search grounding, and portfolio constraints. The Darwinian Loop: Agent influence is dynamically updated based on historical predictive accuracy and PnL attribution. Synthesis & Execution: A Portfolio Manager (PM) Agent synthesizes the debate, applies hard-coded risk limits, and executes a paper trade through the Alpaca API. The result is a dramatic reduction in decision latency from minutes to seconds, while increasing explainability, risk governance, and research coverage.

How we built it The architecture is built on a multi-agent "Darwinian" framework powered by the Gemini API, orchestrated via a FastAPI backend and LangGraph.

1. Evidence Layer & Tool Use We utilize Gemini 1.5 Flash for low-latency, parallel tool-calling. Agents dynamically query Polygon.io for price/options data, scrape news APIs for sentiment, and ground their arguments in factual, real-time data.
2. The Adversarial Debate Specialized agents analyze the evidence in parallel: The Bull Agent: Scans technicals, fundamentals, and product cycles for upside. The Bear Agent: Scans sentiment, short-interest, and supply-chain data for downside. The Risk Officer: Checks current portfolio exposure, correlation, and volatility regimes.
3. The Darwinian Weighting Loop The system doesn't treat all agents equally. We implemented an evolutionary scoring system where agent influence is updated based on historical predictive accuracy (Brier score) and PnL attribution. The agent weight update is then calculated using a softmax function over the accumulated scores to determine capital allocation authority
4. Synthesis & Execution The Portfolio Manager (Gemini 1.5 Pro) synthesizes the debate, applies the dynamic weights, and outputs a structured JSON trade ticket. A deterministic Python Risk Engine validates the ticket against hard limits (e.g., max sector exposure). If passed, the system uses the Alpaca API to execute the trade.

Challenges we ran into Latency in Multi-Agent Debates: Sequential agent calls took too long for event-driven trading (often >45 seconds). Solution: We utilized Gemini Flash for parallel evidence gathering and debate generation, reserving the heavier Gemini Pro model strictly for the final PM synthesis. This reduced decision latency to under 3 seconds. Unstructured LLM Outputs: Financial execution requires strict, unbreakable schemas. A single malformed JSON response could crash the execution pipeline. Solution: We heavily utilized Gemini's native JSON schema enforcement and Pydantic validation to ensure every agent output could be programmatically parsed and executed flawlessly. The "Black Box" Trust Problem: Institutional users cannot trust an AI they cannot audit. Solution: We built a "Trade Auditor" UI that traces every execution back to the exact API calls, agent quotes, and risk checks that led to the decision, providing full explainability and an immutable audit trail in PostgreSQL. Accomplishments that we're proud of Sub-3-Second Execution: Achieved an end-to-end event-to-decision latency of <3 seconds, proving that multi-agent reasoning can operate at the speed of modern markets. True Autonomy: Built a system that doesn't just suggest a trade, but holds the API keys, sizes the position based on Kelly Criterion logic, and pulls the trigger via Alpaca. Early Hackathon Traction: Achieved $4,500 in recognized MRR during the 90-day sprint by onboarding 3 beta proprietary trading desks, proving immediate market demand for our logic streams. Mathematical Governance: Successfully implemented a live Darwinian learning loop where agents are mathematically penalized for poor calibration and rewarded for alpha generation. What we learned Adversarial Reasoning Beats Single-Agent Prompting: A single LLM is prone to confirmation bias. By forcing a "Bull" and "Bear" agent to debate, we drastically improved directional accuracy and eliminated single-model hallucinations. AI Needs Hard Guardrails: You cannot trust an LLM to enforce its own risk limits. We learned that the reasoning engine (Gemini) must be strictly separated from the execution engine (Hard-coded Python risk limits). The AI recommends; the deterministic code enforces. Trust is the Product: In finance, explainability is just as important as accuracy. Building the audit trail was just as critical as building the trading logic. What's next for Origin - Black-Box Arena Expand the Agent Council: Introduce specialized agents for Options Flow, Macro Regimes, and Supply Chain logistics to capture deeper alternative data signals. Enterprise Live Execution: Transition from paper-trading to fully audited, SOC-2 compliant live execution environments for enterprise family offices and hedge funds by Q1 2027. The Agent Marketplace: Build a platform where quantitative researchers can fork, fine-tune, and deploy their own specialized agents into the Arena, creating a decentralized marketplace for synthetic alpha.

Built With

• fastapi
• gemini
• langraph
• pandas
• polygon
• postgresql
• pyhon
• react
• tailwind