AlphaFlow presents

MicroWorld

A world model of finance.

We are building a market where thousands of LLM-driven agents, each standing in for one real participant, play the game under real rules. We start on-chain, where every position is public.

Pre-revenue · Working prototypes on real on-chain data

$19 billion in one day. Nobody saw the chain.

On October 10, 2025, more than $19 billion of crypto positions were liquidated in a single day. Each venue saw just its own book. Risk tools look at past statistics and at one portfolio at a time. None of them see how everyone's forced selling interacts.

Leverage is everywhere.

Perpetual futures now trade tens of trillions of dollars a year.

Everyone sees just their own book.

Venues, lenders and funds each manage risk in isolation.

AI agents are crowding in.

Similar models make similar moves, and crowded moves become cascades.

Built on real positions.

Star Map

Every star is a real open position on Hyperliquid. Replay a crash minute by minute against real historical prices (October 10, 2025) and watch positions cross their liquidation prices. A scenario replay, not a record of that day.

What-If Engine

Choose a shock. The engine checks every position against its liquidation price, aggregates forced selling, moves prices and repeats until the market settles.

~3,100 accounts · ~8,800 positions · ~$7.8B notional · about a third of open interest

BTC −8% in 15 minutes under stressed liquidity → −9.7% becomes about −11.7% once forced liquidations feed back.

Simulation on real positions with literature-standard parameters. An illustration, not a forecast.

Markets are a game. We let AI agents play it.

Every participant (a whale, a market maker, a 20x-leverage trader, a lending vault) will be played by an LLM-driven agent that stands in for one real participant, with its real positions, leverage and limits. The agents trade, hedge, panic and get liquidated against each other, under the venue's real margin and liquidation rules. Prices and crashes emerge from their game instead of being assumed.

Stage 1: Web3, where the game is visible.

On-chain exchanges publish every account's positions, leverage and liquidation price. We reverse-engineer the players: we reconstruct each account's state, learn how it actually behaves from its on-chain history, and check simulated agents against what really happened. Web3 is our wind tunnel. Near-term product: neutral, cross-venue risk infrastructure for venues, lenders and stablecoin issuers.

Stage 2: all of finance, where the game is hidden.

On-chain markets are still a small slice of global finance. In traditional markets positions are private, but the rules of the game are similar, and the method tested in Stage 1 carries over. Participants are inferred from public filings, mandates and aggregate market data, then calibrated the same way.

Every important financial decision should be simulated before the market sees it.

Team

Building risk infrastructure for on-chain markets? Let's talk.

alphaflow.fintech@outlook.com