Yanyu: studying AlphaPai with REA and independently rebuilding research workflows with Codex

[lxistired]

Hi REA community — sharing Yanyu (研屿), an independent open-source case study in using REA + Codex to study AlphaPai and rebuild selected financial-research workflows.

We used REA for static analysis of parts of the AlphaPai client, then Codex to independently implement related workflows. This is a bounded reimplementation of selected workflows; it does not reproduce the original backend. We do not distribute original AlphaPai source code or data, and this project does not imply endorsement by REA or AlphaPai.

The goal is to help securities firms, online brokerages, and independent research institutions build and control their own research tools and customer-service workflows. We provide open-source tools and integration interfaces, not inference services. Institutions are responsible for choosing, procuring, deploying, and paying for model APIs or local models, and for their own data governance.

The preview covers implemented or simulated paths for uploads and search, PDF documents, references, research plans, and meeting/recording workflows. Implementation and validation differ by feature:

• Published validation records cover core component regressions in a simulated environment, type checking, the build, and a real local static-HTTP service check.
• Summaries and general automation are explicitly simulated. PDF search works on extracted text; it does not provide original-page coordinate mapping or OCR.
• Real-browser interaction, real devices, external services, and visual parity with the original product remain unvalidated. This is an early preview, with known gaps documented in the repository.

Links:

• Source, current documentation, and Linux startup instructions
• First public preview release: source and matching build
• Validation scope and known gaps

Our own code is MIT-licensed. Thanks to the REA maintainers and contributors for the analysis tools. Feedback on the analysis-to-implementation workflow, reproducibility, and validation boundaries would be welcome.