Generative AI Hackathon with IBM Granite

Start

Validly

Christopher Little-Savage

Joe O'Connor

KerryV-R

70% of the largest companies in the world run on legacy infrastructure. They spend billions of dollars each year trying to modernise their codebases, in a process that is often manual, slow and expensive (projects may cost $2M+ and take 6-16 months to complete).

It’s possible to transpile COBOL to Java using off-the-shelf tools which preserve semantics but produce illegible, unmaintainable code. Our innovation is the ability to use AI to guide the incremental application of small, provably-correct code transformations to make the code readable and maintainable without affecting its execution semantics. We generate these transformations with an LLM, but then prove them correct ahead of time, gradually building a huge repository of validated transforms. We’ll can sell this as a service to existing consultancies that work in the space or work with clients directly ourselves.

Over time, we can apply the same approach to new use cases. We could help companies move from C to Rust to comply with CISA cybersecurity recommendations. Alternatively, we can use these LLM-guided techniques within a language like CUDA to improve performance rather than maintainability. In the general case, our technology can be applied to refactor and improve any code with correctness guarantees not provided by existing LLM methods, presenting opportunities for better AI-generated code in any domain.

Validly combines the creativity of LLMs with formal methods from computer science to safely rewrite code automatically. To start with, we’re tackling legacy code modernization from COBOL to Java. 

"Validly team on Generative AI Hackathon with IBM Granite Hackathon"

Team Idea

Submission