Project archive
SYS-01 / MLOps · drift monitoring

DriftPainel

Monitor drift between training data and current data to decide when to observe, keep or retrain models.

STATUSPublished on GitHub
ROLEFull-stack architecture, MLOps and technical documentation
FLOW / SYS-01
  1. 01Dataset
  2. 02Perfil
  3. 03Testes
  4. 04Score
  5. 05Worker
  6. 06Relatórios
  7. 07Dashboard
EXECUTIVE SUMMARY

A full-stack MLOps platform that receives datasets, calculates feature-level drift severity, ranks relevant changes and generates HTML, Markdown and JSON reports. The FastAPI backend orchestrates rules, Celery runs analyses and the React dashboard turns signals into operational decisions.

  • FastAPI
  • React
  • Celery
  • Redis
  • PostgreSQL
  • MinIO
  • Docker
SYS-01.01

Context and problem #

Production models lose reliability when data distributions change. The project turns that uncertainty into an observable routine with severity and history.

SYS-01.02

My contribution #

The case demonstrates asynchronous backend work, dashboarding, persistence, file storage and technical decision documentation in an integrated platform.

SYS-01.03

Architecture #

FastAPI receives uploads and triggers services; Celery runs the drift engine; PostgreSQL stores metadata; MinIO or local storage keeps files and reports; React exposes history, charts and downloads.

  • FastAPI REST
  • Celery + Redis
  • PostgreSQL + MinIO
  • React dashboard
SYS-01.04

Evaluation and quality #

The value is making drift measurable per feature and run, enabling comparison across global severity, rankings and action recommendations.

SYS-01.05

Limitations #

Drift indicates statistical change, not causal model degradation. Retraining decisions still need business impact and model metrics.

SYS-01.06

Next steps #

Evolve authentication, multi-user support, Prometheus/Grafana observability, scheduling and MLflow integration.