P.O.O.K.I.E — Portfolio Optimization Opportunity Knowledge Intelligence Engine

Inspiration

Prediction markets are the ultimate test of collective intelligence — yet they’re dominated by institutional traders with data, algorithms, and capital. Retail traders, on the other hand, rely on gut feeling and hype.
We wanted to break that wall. What if anyone could run their own quantitative hedge fund, powered by AI?

P.O.O.K.I.E was born from that question — and from the belief that AI can make markets smarter, fairer, and more transparent.

What it does

P.O.O.K.I.E (Portfolio Optimization Opportunity Knowledge Intelligence Engine) is an autonomous AI trading system that finds and exploits inefficiencies in prediction markets like Kalshi.

It performs unbiased research using Perplexity AI, translates insights into structured probabilities with Google Gemini 2.5 Pro, and applies quantitative finance logic (R-scores, Kelly Criterion) to detect statistically mispriced markets.

It then generates transparent, data-backed trade recommendations — displayed in a sleek dashboard with live confidence metrics, portfolio tracking, and AI reasoning.

How we built it

We built P.O.O.K.I.E as a two-part system:

Backend (Python) handles market discovery, AI research orchestration, and trading logic. It integrates the Kalshi API, Perplexity Sonar for independent research, and Google Gemini 2.5 Pro for structured decision outputs.
Frontend (Vite + React + TypeScript) visualizes those results in real time. Using shadcn/ui and Tailwind CSS, we built a production-grade dashboard with panels for suggested bets, investment theses, and portfolio tracking.
AI Pipeline connects everything: Perplexity → Gemini 2.5 Pro → Pydantic validation → JSON export → Dashboard rendering.

The result is a seamless, research-to-decision pipeline that feels like having an AI quant team on your laptop.

Challenges we ran into

Kalshi authentication required RSA signatures — we implemented custom cryptographic signing with timestamp handling.
AI bias: Early models “peeked” at market prices and anchored their predictions. We fixed this with a two-phase research process: first research without odds, then combine results later.
Structured output reliability: We initially struggled with parsing errors from other models, so we switched to Gemini 2.5 Pro’s JSON mode for near-perfect success rates.
Performance: Sequential API calls were too slow; async batch processing cut runtime from minutes to seconds.

Accomplishments that we're proud of

Built a real, production-ready trading engine, not a demo or wrapper.
Achieved 100% structured output success rate using Gemini 2.5 Pro’s JSON mode.
Designed a clean, accessible dashboard with professional UI/UX and real-time updates.
Created a fully autonomous system that can analyze, reason, and act without human input — breaking conventional barriers in both AI and finance.

What we learned

We learned that AI is only as powerful as its structure — orchestration between models matters more than raw intelligence.
We also discovered how critical unbiased research is for reliable predictions, and how statistical reasoning can bring accountability to AI decision-making.
Finally, we learned how to merge AI engineering, quantitative finance, and modern frontend design into one cohesive product.