Skip to content
All projects01 / DRIFT-AWARE-MLOPS
Bachelor Project

Drift-Aware MLOps Pipeline

A lightweight, configuration-driven MLOps pipeline for governed model lifecycle management under changing data. Traceable training, evaluation, promotion, rollback, local serving, monitoring, and drift-adaptive retraining across tabular, image, and RAW-image workflows.

ROLE
Systems Architect / ML Engineer
CONTEXT
Academic / Independent
DURATION
5 months
YEAR
2025–2026

Interactive pipeline run

// LIVE · PIPELINE RUN
EXEC

PIPELINE.run#4172
IDLE
  • 01preprocessWAIT
  • 02trainWAIT
  • 03evaluateWAIT
  • 04drift analysisWAIT
  • 05promotion gateWAIT
  • 06human approvalWAIT
  • 07deployWAIT
  • 08monitorWAIT
DRIFT · PSI—
0.00thr 0.150.30
CAND ACC
—
PROD ACC
0.92
PROMOTION.gatePENDING

Awaiting evaluation

0.0s

An interactive model of the real lifecycle. Press play to run it, watch drift cross its threshold, and make the promotion call, the one the pipeline gates on a human.

Problem

ML systems get hard to maintain once data versions, preprocessing, training runs, evaluation, registry state, and monitoring live as disconnected artefacts. When production data shifts, teams need a reproducible way to trace lineage, detect drift, and decide whether to retrain, promote, or roll back, without enterprise-scale infrastructure.

Solution

A local-first, batch-oriented framework that turns the model lifecycle into an auditable, governed workflow. It connects dataset versioning, preprocessing, training, evaluation, drift analysis, MLflow metadata, rule-based promotion checks, human approval, rollback records, deployment manifests, and FastAPI serving. Drift is decision support for investigation or retraining, not a blind automatic trigger.

Highlights

  • End-to-end lifecycle: preprocessing, training, evaluation, promotion, deployment manifests, monitoring, and drift-adaptive retraining
  • Rule-based promotion gates: candidates checked against quality thresholds, compared to the current Production model, enriched with drift evidence, and gated by explicit human approval
  • Drift as a decision signal: Evidently for tabular, plus custom pixel, embedding, and multi-scale analysis for images
  • Full traceability across dataset version IDs, config hashes, run IDs, seeds, and MLflow metadata; rollback as a governed registry transition
  • Core tabular pipeline verified in GitHub Actions: tests, training, evaluation, Docker build, and container smoke test
LIFECYCLE STAGES
8
DATA MODALITIES
3
DRIFT METHODS
4