LLM

Jump to TL;DR We’ve now reached a point where the platform can take a Java file, run it through the full workflow, and return a refactored version with no manual steps in between. That’s thanks to two new agents, coverage integration, and automated file handling, all stitched together into one continuous pipeline. 1. Code Reviewer Agent The Code Reviewer is our new quality gate between transformations, and is now fully implemented. When the Code Transformer makes changes, the Reviewer checks them for correctness, quality, and whether they actually match the intended refactoring strategy. ...

Jump to TL;DR The last week of July has seen us work extensively on our agentic workflow and tooling support. We’ve also been researching how to best implement prompts in our solution to decouple them from the agents and allow easy modification and insertion of custom prompts by our users. We are also instrumenting our Anti-Pattern Remediator to enable observability, debugging, and evaluation of the agents. 1. Prompt Templates Maoqin performed an extensive research spike to find the optimal way of separating prompts from our agents to facilitate plug-and-play, while also increasing our AI agents’ capabilities. ...

Jump to TL;DR We have focused this week on important architectural improvements and clarifications, and pushing toward a refined solution by working on new agents and support tooling. 1. Architecture Context Diagram Container & Component Diagram Our architecture has become clearer in our mind. Our project’s official title is now Anti-Pattern Resolutor, and comprises multiple projects: ...

Jump to TL;DR This week, we had the opportunity to present our MSc group project at the Exploring AI Proof of Concepts with UCL and IBM event, hosted at IBM Innovation Studio London and sponsored by IBM TechXchange. We were joined by fellow UCL students, IBM engineers, and industry stakeholders, and we’re grateful for the warm welcome from the IBM team, our mentors, and the wider community. 1. A Day of Inspiring Presentations The event brought together students from across UCL to showcase a wide range of AI-powered proof-of-concept solutions. Some were developed as part of group projects, others independent research. It was incredibly rewarding to see such thoughtful, technically ambitious work addressing real-world challenges in fields like Healthcare, Sustainability, and Education. ...

Jump to TL;DR Following last week’s groundwork in static analysis and semantic retrieval, our focus this week has shifted toward multi-model orchestration, LLM infrastructure, and preparing for our upcoming IBM showcase on the 16th. 1. Unified Provider Interface To enable flexible experimentation and future-proofing, we’ve implemented a unified interface for interacting with multiple LLM providers. Our abstraction currently supports Ollama, vLLM, and WatsonX, but, following the Open-Closed Principle, we’ve made sure to facilitate the addition of new ones with no friction. ...