Lume – Architecting an Agentic Prediction Market

Built a prediction market from scratch at Eternis.ai, where AI agents and humans trade on real-world outcomes.

Role: Founding Product Designer + Design Engineer

Timeline: Ongoing (2025–Present)

Company: Eternis.ai

By Bishal Mishra, Product Designer

Architecting an Agentic Prediction Market

Built a prediction market from scratch, where AI agents and humans could trade.

Overview

Lume is a prediction market built at Eternis. Users bet on real-world outcomes through an opinionated interface that makes complex financial mechanics feel intuitive and actionable. I led the product end-to-end: deciding what to build, designing the experience, and engineering the frontend.

Outcomes

Scoped, designed, and engineered the full product from zero: betting flows, market discovery, AI agent trading, leaderboards, and viral growth loops.

Challenges

Prediction markets are inherently complex. Multiple outcomes, shifting odds, real stakes. The goal was to reduce all of that to something anyone could act on within seconds, without losing the depth that keeps experienced traders engaged.

Familiarity

The interfaces assume familiarity with order books, liquidity, and odds formats. That's a ceiling on adoption. I mapped the competitive landscape and used AI to pressure-test early hypotheses, arriving at a tighter product direction before committing to any design work.

Agents trading first, humans second

The product launched with AI models like ChatGPT, Claude, and Kimi trading against each other on real markets. Users could watch, follow, and track how each model performed. Once the mechanics were proven, we opened it up to human traders. That sequencing shaped every product decision that followed.

Yes 36¢. No 67¢.

I reduced the betting experience to cent-based pricing. Order books are hidden at first glance, no percentage confusion. A user sees 36 cents, immediately understands the upside. Each market card communicates the question, yes/no prices, potential payout, deadline, and volume in one glance. No click required to understand the bet.

Picking a side to bet

The betting page frames every market as a debate. Real figures take opposing positions with actual quotes, making the bet feel like a stance rather than a transaction. The cent-based pricing, visual odds split, and live holder leaderboard all sit on one page. No tabs, no hidden complexity. Everything a user needs to decide is visible the moment they land.

Distribution

Every bet a user places is inherently shareable. I designed a Twitter-native sharing experience that turns each bet into a rich card with the market question, the user's position, and current odds. Users bet, share, and their followers click through. The product distributes itself.

Anyone can create a market. AI decides if it's valid

Users propose any market they want. AI validates whether the question is resolvable, well-scoped, and fair. Once approved, the market goes live and funds its own liquidity. This removes the editorial bottleneck without sacrificing quality.

Use of AI

I used Midjourney and Weave to generate distinct illustrations throughout the platform, giving a unique visual identity. On the engineering side, I used Claude Code to prototype faster and push code to production.

There's more to this story.

This case study covers the key decisions, but there's a lot more behind the product: iterations that didn't make it, technical trade-offs, growth experiments, and lessons from shipping fast with a small team. Happy to walk through any of it in conversation