SODAOpt: a transformer framework for smart task assignment in software teams

Imagine a large IT company with thousands of tasks and hundreds of developers. How do you automatically assign the right person to the right task, taking into account not only past experience but also context — the task description, commits, role and profile of the developer?

This is exactly the problem addressed by SODAOpt, our transformer-based framework for intelligent task assignment in software development.

What is wrong with existing approaches? Most systems rely on static identifiers (a developer ID) and historical statistics, and completely ignore both the semantics of the task and the professional profile of the specialist.

How SODAOpt works:

• a text encoder based on E5-large extracts meaning from task descriptions and commits;

• adaptive layers combine textual features with the developer profile;

• a composite loss function combines contrastive learning with assignment optimisation;

• the optimal match is retrieved through a FAISS index.

In other words, the model "reads" the task description while simultaneously studying the developer's profile: their experience, role and department. It then merges the two streams of information and finds the best match — the person on the team the task suits most. The more context, the more accurate the assignment.

The model substantially outperforms competing ID-only and text-only approaches.

The paper was published at ACM FSE Companion '25 (Trondheim, Norway). Code: github.com/SODAForge/SODAOpt