🚀 Inspiration

Prediction markets already surface brilliant signals, but we kept losing the thread between spotting a mispricing, proving it’s incoherent, logging the alert, and prepping the resolution receipts. Kairos was born to be that “markets mission control”—something a solo trader or a newsroom desk can open and instantly see how AI agents, watchlists, math, and evidence fit together without tab chaos.

🛠️ How I Built It

Stack:

• Next.js 15 App Router + React 19, Server Components, Server Actions, Tailwind v4, Radix UI, Recharts for sparklines, and Lucide icons.
• Authentication rides NextAuth 5 beta with the Prisma adapter and a Neon/Postgres database (JSONB fields store Claude memory + Polymarket credentials).

Data layer:

• Custom clients for Polymarket Gamma + CLOB, Nevua, Adjacent, and Anthropic Claude. Each helper normalizes env vars, throws typed errors, and ships demo fallbacks so the UI keeps working offline (hackathon-friendly).

APIs & caching:

• /api/polymarket/price-history memoizes token/interval pairs for 30 s; Nevua and Adjacent fetchers run on the server with concurrency caps (MAX 8 news lookups) to stay within rate limits.

Math & logic:

• Coherence encodes overlapping markets as a binary matrix (rows = markets, cols = world states) and checks whether price vectors violate ∑p_i = 1. Probability gaps generate trade recommendations with limit prices and certificates.
• Factor runs belief propagation on a six-node graph and exposes attribution by storing deltas per edge, while a knapsack DP maximizes signal lift under a toggle budget.
• Arena treats loophole coverage as a set-cover problem; we mark severity, patch type, and residual risk so judges see why the spec is now airtight.

Agent memory:

• Claude conversations persist via Prisma claudeMemory JSONB, exposed through /api/memory, so every module (alerts, factor, resolve) can read/write the same state.

UI polish:

• Motion-powered reveals, shader backgrounds, responsive side nav, and turbopack dev for instant feedback.

🧠 What I Learned

• React Server Components + Server Actions drastically simplify multi-API products when you keep all secrets on the server.
• Traders trust automation only when they see the math; even lightweight linear algebra + DP visualizations go a long way.
• Good fallbacks (seed data, graceful errors) are essential for hackathons where API quotas will fail right before judging.
• Prompt engineering is easier when the rest of the stack has typed schemas (Nevua alert JSON, watchlist payloads, etc.).

🧗 Challenges I Faced

• Wiring four external APIs with different auth styles while keeping keys server-only on a React 19 app router.
• Handling Polymarket CORS/timeouts; we solved it with our own caching route + seeded fallback data.
• Getting NextAuth 5 beta + Prisma to play nicely with the new React Server Components data flow.
• Translating qualitative prompts (“ping me if price jumps 5%”) into strict Nevua alert JSON without exposing secrets.
• Designing math explanations that feel trustworthy to market makers as well as policy analysts.

✨ What’s Next

• Connect to Polymarket’s order API so trade vectors can be executed (with risk rails) straight from Coherence.
• Stream live news + price deltas via websockets instead of refetching.
• That’s the full story behind Kairos—logic, math, and all the glue that keeps prediction-market operators in sync.

A HUGE SHOUTOUT TO NUEVA AND ADJACENT FOR PROVIDING ME PERSONALLY WITH THEIR APIs !