My master’s thesis: an AI assistant that turns feature descriptions into implementation plans
ATLAS (Assisted Task & Logical Architecture Support) is my master’s thesis, built at Tieto. It helps developers go from a feature description to a concrete plan: which tasks to do, in what order, and where in the codebase to start.
I designed and built the prototype. You upload the feature documents, and ATLAS answers with a short summary, sequence and dependency diagrams, and grouped tasks that point to the files involved.
Behind it are four services, as in the diagram below. The UI sends the request to a chat backend, which queues it and asks a retrieval backend for context. The retrieval backend searches the knowledge base and returns the most relevant pieces. The chat backend then sends the request, that context and a system prompt to the LLM backend, stores the answer and returns it to the UI.
The heart of it is retrieval (RAG). The knowledge base holds the codebase as code, documentation and AI-written module summaries, so ATLAS can explain the architecture and still point to real files. Each answer gets a small, varied selection of that context, and the model may only name files that retrieval actually found.
The code belongs to Tieto. The full thesis is linked below.
Built with: Python, JavaScript, LlamaIndex, OpenAI API, RAG, SQLite, Mermaid, Figma
