Already Done
CME changed its feed. By the time an engineer looked, the fix, the test and the pull request were already waiting.
An open-weight model watches CME's public market data schema and channel configuration, decides whether a feed handler is affected, writes the change and opens a tested pull request. It is replayed against CME's real history, from schema version 9 to today's channel file, and scored against an answer key written before it ran.
How this was built
Every exchange changes its market data feed several times a year: new channels, products moved between channels, new status values, new schema versions, new multicast addresses. Someone at every data vendor reads each notice, works out whether it matters, changes the feed handler, tests it in the exchange's certification environment and ships it before the effective date. Across hundreds of venues, that's a permanent team, and most of the work is reading, diffing and careful editing.
This lab hands that job to a model. A small reference feed handler for CME's MDP 3.0 feed, written in Rust, records exactly which CME schema and channel file it is certified against. When CME publishes a new one, plain code moves that pin, regenerates test packets from the public schema and runs the checks. The model reads what failed, classifies the change, edits the code until the checks pass, and writes the pull request, including what it wasn't sure about. A person approves every change; the model can't push, merge or deploy.
To show how often it gets this right, the system is replayed against CME's real history, one file at a time, and scored against an answer key committed to the repository before the first step ran. Everything runs on open-weight models in Amazon Bedrock: GLM-5 for triage and the code, and Kimi K2.5 as an independent reviewer. The models were chosen by trying candidates on real CME changes. No market data is used: test packets are generated from CME's public schema.
Exploring AI for market data feed maintenance, November 2026. Runs on GLM-5 in Amazon Bedrock.
A personal project, built on a personal AWS account using personal equipment. Not affiliated with my employer.
This demo is interactive and needs JavaScript.