AutoAMM
An autonomous research lab that tests and refines programmable market-making strategies.
Demo
About
AutoAMM brings the autoresearch loop to market making: propose a strategy change, run an experiment, evaluate it against the incumbent, and keep or reject the result. The prototype searches fee levels and active capital allocations for a WETH/USDC constant-product strategy. Candidates must pass drawdown and inventory limits and improve on both training and validation scenarios. A separate holdout is unlocked at the end of a 24-experiment run. The Next.js dashboard exposes the experiment journal, rejection reasons, performance comparisons, strategy code, and downloadable evidence. Accepted parameters compile into real 1inch Aqua/SwapVM program bytecode and unsigned transaction plans. The public demo uses a deterministic heuristic optimizer and synthetic order flow; it does not sign transactions or claim simulated results are live trading profits. A server-side Graph adapter supports indexed market prices when credentials are configured.
How it's made
Built with Next.js, React, TypeScript, Recharts, Zod, viem, and the official 1inch Aqua and SwapVM SDKs. The deterministic constant-product simulator models fee-sensitive order flow, arbitrage, retained fees, and estimated execution costs. An autonomous search loop explores fee levels and capital allocations while a fixed evaluator checks net equity relative to HODL on training and validation paths. The dashboard streams the resulting experiment history and unlocks a separate holdout after 24 candidates. The Aqua compiler combines flatFeeAmountInXD with xycSwapXD and prepares token approvals and ship calldata using the official SDKs. A server-side Graph adapter queries and validates hourly WETH/USDC pool prices. The Vercel deployment intentionally contains no market-data, model, or wallet keys. Simulation and bytecode compilation work without credentials; live-data access and transaction signing are separate. Drawdown and inventory-percentage gates are offchain evaluator checks, not guarantees encoded in the compiled program. Codex provided substantial coding and design assistance.