Async classification
FastAPI plus RabbitMQ runs single and batch jobs on a separate inference service, with WebSocket updates for live progress.
Back
University Project
ML-Powered Detection of Architectural Design Decisions in Jira Issues
An end-to-end system that detects ADDs inside Jira issues and exposes them through a fast asynchronous API and a responsive React UI, backed by full MLOps and observability.
What it is
A production-shaped ML pipeline for ADD detection.
Architectural design decisions are scattered across Jira issues and easy to lose. The system trains a lightweight BERT classifier to surface them, then wires the model into a queue-based execution path so single and batch classification stay responsive under load.
The platform layers MLOps and observability on top: DVC and Great Expectations version and validate the data, MLflow tracks model artifacts and metrics, and Prometheus, Grafana, and Loki cover latency, traffic, and logs. GitLab CI/CD ties build, validation, training, testing, and deploy into a single repeatable flow.
Versioned data and models
DVC and Great Expectations validate inputs, MLflow tracks model versions and metrics, and Postgres stores issues with keyword search.
CI/CD as the backbone
GitLab pipelines cover build, data artifacts, validation, training, testing, and deploy on a VM, with secret management baked in.