I Built an AI That Modernizes Legacy COBOL — And Then Made It Prove Its Own Work

2026-04-02 · Source: LLM on Medium · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Advanced, long

Summary

A new AI-driven system, Py-BOLD, addresses the $3 trillion problem of modernizing legacy COBOL code, which processes 95% of ATM and 80% of credit card transactions daily. Traditional AI migration tools often produce "Py-BOL"—Python code that retains COBOL's opaque naming conventions, making it unmaintainable. Py-BOLD separates the tasks of preserving correctness and restoring business intent. It uses a deterministic compiler layer to generate a typed Abstract Syntax Tree (AST) and Data Flow Graph (DFG) from COBOL, ensuring logic integrity. This AST then feeds into a LangGraph agent pipeline, where a Structural Agent performs a mechanical COBOL-to-Python translation, a Semantic Agent renames variables to reflect business intent, and a Verification Agent automatically generates pytest suites to confirm behavioral parity with the original COBOL. This neuro-symbolic approach prevents LLM hallucinations from corrupting core business logic.

Key takeaway

For AI Architects and Machine Learning Engineers tasked with modernizing legacy systems, consider adopting a neuro-symbolic architecture like Py-BOLD. By strictly separating deterministic logic preservation via compiler theory from probabilistic semantic interpretation by LLMs, you can prevent silent logic corruption and ensure verifiable code transformations. Focus on building robust ASTs as the ground truth and use LLMs only for constrained, auditable tasks like renaming and test generation to maintain system integrity.

Key insights

Separating correctness (compiler theory) from intent (LLM semantics) is key to reliable legacy code modernization.

Principles

• Deterministic parsing anchors LLM outputs.
• Constrain LLM to prevent logic hallucination.
• Automated verification ensures behavioral parity.

Method

Py-BOLD uses a deterministic compiler to create an AST from COBOL, then an LLM agent pipeline (Structural, Semantic, Verification) translates to Python, renames for clarity, and generates tests to prove correctness.

In practice

• Use AST as the single source of truth for LLMs.
• Implement Python-to-Python semantic refactoring.
• Generate pytest suites from AST for verification.

Topics

• COBOL Modernization
• Neuro-Symbolic AI
• Legacy Code Migration
• Abstract Syntax Tree
• LangGraph Agent Pipeline

Code references

• Y-Watanabe11/Py-BOLD

Best for: AI Engineer, AI Architect, Machine Learning Engineer

Related on AIssential

• Reimagining API modernization with deterministic AI-assisted engineering — AI-assisted, instruction-driven frameworks accelerate legacy API modernization by codifying patterns and ensuring deterministic outcomes.
• FM-Agent: Scaling Formal Methods to Large Systems via LLM-Based Hoare-Style Reasoning — LLMs can automate formal specification generation and Hoare-style reasoning for large-scale software systems.
• Learnings from COBOL modernization in the real world — Successful mainframe modernization with AI requires deterministic reverse engineering to provide bounded, platform-aware context for forward engineering.
• Neurosymbolic Repo-level Code Localization — Neurosymbolic AI combining LLMs and Datalog overcomes keyword shortcuts for precise code localization.
• Interpretable and Verifiable Hardware Generation with LLM-Driven Stepwise Refinement — LLMs can generate verifiable hardware designs by integrating formal methods and stepwise refinement.
• Natural Synthesis: Outperforming Reactive Synthesis Tools with Large Reasoning Models — Neuro-symbolic AI can significantly advance reactive synthesis by combining large reasoning models with formal verification.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by LLM on Medium.