Projected 50% less manual work
An agent that reads legacy code in a proprietary language, documents it, and extracts the business rules inside.
Read as steps
- Legacy scripts written in the proprietary language come in.
- A LangGraph dual-state dynamic workflow coordinates the work.
- It calls vLLM-served models, with retrieval finding the relevant context.
- PII filtering keeps personal information out.
- EBNF grammars constrain every output, so it is always well-formed.
- What comes out: documentation, extracted business rules and code prepared for migration.
Context
A client’s business rules live in legacy scripts written in a proprietary programming language. The scripts need maintaining and documenting now, and migrating later. Because the language is proprietary, general-purpose models can’t be trusted with it out of the box.
What I did
I lead the agent’s design and delivery. A LangGraph dual-state dynamic workflow coordinates vLLM-served models with retrieval. EBNF grammars constrain every output to what the grammar allows, and PII filtering keeps personal information out. The agent writes documentation, extracts business rules, and prepares the code for migration.
Outcome
Projected 50% less manual work. In progress.
Evidence
Client work, no public page.
Stack
Python, LangGraph, vLLM, RAG, EBNF structured outputs, PII filtering