STRATEX: Autonomous AI Agent

Inspiration

Prediction markets like Polymarket are incredibly efficient, but they rely on human traders manually reading news, synthesizing sentiment, and pricing risk. When we looked at existing "AI Agents" being built for trading, they were little more than wrappers executing random logic inside a black box.

We built STRATEX to answer a specific hypothesis: Can an Autonomous AI ingest raw orderflow, synthesize real-time news faster than a human, and express its confidence mathematically using professional quant sizing models?

What it does

STRATEX is a fully autonomous, live-data trading agent designed for the AI-Augmented Systems track. It operates without human intervention but features a "Command & Control" (C2) dashboard for manual override.

Instead of hardcoding algorithmic backtests, STRATEX generates edge via Information Velocity.

The agent fetches a live market question from the Synthesis API (e.g., "Will TikTok be banned?").
It queries a live web search to scrape the top 3 most recent articles on the topic.
It passes this context to Gemini 2.5 Flash, demanding a True Probability $p$ evaluation.
If the AI's $p$ differs from the market's implied probability $m$ by greater than 5%, it recognizes an Expected Value (EV) edge and executes a simulated paper trade.

Architecture & Data Sources

We built a clean, threaded Python backend that operates across four distinct layers:

Data Layer: 100% powered by the Synthesis API for real-time orderflow (Targeting the $200 prize).
Inference Engine: Google's Gemini 2.5 Flash for rapid probability synthesis combined with real-time duckduckgo-search.
Execution & Ledger: Custom Python engine supporting threading, lock-safe caching, and a paper-trading portfolio.
Frontend UI: A custom-built Dash application rendered as a "Cyber-Quant Terminal," complete with a live scrolling thought-process log and interactive C2 manual override mechanisms.

Strategy Logic & Quant Sizing

We didn't want the agent guessing flat dollar amounts. To prove our math, we implemented the Half-Kelly Criterion.

$$ f^* = p - \frac{1 - p}{b} $$ (Where $f^$ is the fraction of the bankroll to wager, $p$ is the AI's probability of winning, and $b$ is the net odds received on the wager).*

The agent dynamically calculates $f^$, scaling its capital aggressively on massive inefficiencies, and minimizing exposure on weak edges. It then logs the exact Expected Value (+EV) of the trade to the portfolio ledger. To protect capital from algorithmic hallucination, we enforced an automated 20% Drawdown Circuit Breaker.

Performance Metrics & Measurable Output

During our simulated runs on live Polymarket data, STRATEX achieved:

Consistent Output Generation: An average 10-market scan cycle completes in under 60 seconds.
Measurable Position Sizing: Kelly allocations dynamically shifting correctly between 0% and our 25% safety cap based strictly on Gemini's delta to the market.
Active EV Capture: Visual proof of +EV edge calculation logging live to the frontend ledger.

Challenges we ran into

Integrating the complex Kelly Criterion math across prediction markets (which price YES/NO differently than traditional sportsbooks) required careful algebraic normalization of the odds $b$. Furthermore, ensuring the Dash UI safely polled the background threading loop without crashing required implementing robust threading.Lock() states on the shared data cache.

Accomplishments that we're proud of

Converting a theoretical quantitative concept into a beautiful, functional dashboard. The moment the agent autonomously read a live news article, correctly identified a mispriced Polymarket line, and printed the Kelly mathematics to the terminal was an incredible feeling.

What we learned

We learned that LLMs (like Gemini 2.5 Flash) are surprisingly adept at producing rigorous probability scores when constrained by strict JSON schema prompting and grounded with real-time search context.

What's next for STRATEX: Autonomous AI Agent

The next step is hooking the paper-execution ledger into a real EVM wallet to place live Polymarket execution orders via smart contracts, crossing the bridge from our current high-fidelity simulation to automated on-chain arbitrage.