CodeCraft

Problem

AI coding tools were scattered across separate products, each tied to one model. IBM asked our team for one web app that could analyse, generate, complete and translate code, let the user pick the model best suited to the task, and remember the conversation.

Approach

• Multi-model by design. Different models were better at different tasks, so the UI offers a task and a model (watsonx.ai, GPT-3.5 or Code Llama) with sensible defaults, and LangChain routes the request.
• One submission page that takes pasted code or uploaded files, in several languages, with guidance on what each task does.
• Accounts and memory. Sign-up and login with hashed passwords, conversation history, and a Milvus vector store so the assistant can recall earlier exchanges.
• Shipped like a product. Four two- to four-week sprints on a GitHub Kanban board, Docker Compose for the frontend and backend, Nginx in front, and GitHub Actions for CI.

My part

I led the frontend team and, as the project grew, took on the styling workstream. I worked on the React component structure and the main code-submission experience (text and file input, task and model selection, presenting responses), refined layouts with CSS and Material UI, and agreed request and response shapes with the backend and AI developers. On the delivery side I ran frontend check-ins, delegated tasks during sprint planning and helped prepare the showcase demo.

We were mentored by IBM engineers Mihai Criveti, Ronan Dalton and Panpan Lin. Panpan went on to mentor me again during my IBM internship. This was a team effort of eleven people. The original repository has the full commit history.