Production ML Platform for Predictive Maintenance

Mohammad Ahmadian — AI / Machine Learning Engineer

End-to-end production-oriented ML system for machine failure risk scoring: validated data → leakage-safe features → baseline + LightGBM → evaluation → FastAPI → Docker.

Hiring managers: start with the Demo Website · Performance Dashboard · Architecture · Presentation pack

Recruiter snapshot

| Item | Detail | |------|--------| | Problem | Binary failure risk on industrial sensor features | | Dataset | AI4I 2020 (UCI) · CC BY 4.0 · synthetic benchmark | | Integrity | Failure-mode flags excluded · chronological UDI split | | Models | Logistic Regression baseline · LightGBM primary | | Serving | FastAPI /v1/predict + OpenAPI | | Docs | Full pack under docs/ |

Held-out test metrics

| Model | Precision | Recall | F1 | ROC-AUC | PR-AUC | |-------|----------:|-------:|---:|--------:|------:| | Baseline (LogReg) | 0.571 | 0.138 | 0.222 | 0.901 | 0.323 | | Primary (LightGBM) | 0.952 | 0.690 | 0.800 | 0.944 | 0.781 |

Full report: reports/EVALUATION_REPORT.md

Baseline vs Primary F1

Architecture

System architecture (Released)

Figure 1. System architecture (Released).

Demo & presentation

| Asset | Link | |-------|------| | Demo website | presentation/demo-website | | Performance dashboard | presentation/dashboard | | 90s video script + recorder slides | presentation/video | | LinkedIn / Upwork copy | presentation/copy |

To record the 90s silent demo: open presentation/video/recorder-slides.html full screen → screen-record 1080p → save as presentation/video/demo-90s.mp4.

Quick start

python -m venv .venv
# Windows
.venv\Scripts\activate
pip install -r requirements.txt

set PYTHONPATH=src
# optional retrain:
python -m predmaint.training.train

uvicorn predmaint.api.main:app --app-dir src --host 0.0.0.0 --port 8000
# http://localhost:8000/docs

Sample request: reports/sample_predict_payload.json

curl -s -X POST http://localhost:8000/v1/predict ^
  -H "Content-Type: application/json" ^
  -d @reports/sample_predict_payload.json

Repository map

src/predmaint/     data · features · models · training · evaluation · inference · api
configs/           default.yaml
docs/              PRD → Cursor specs (Released)
presentation/      demo website · dashboard · architecture · video · hiring copy
models/artifacts/production/   served model alias
reports/           metrics · figures · evaluation report

Limitations (honest)

AI4I is synthetic, not a live plant stream
UDI is a sequence proxy, not a wall-clock timestamp
Rare failures → F1 / PR-AUC matter more than accuracy

Contact

mohammad.ahmadian.dev@gmail.com · github.com/ahmadian-dev · Turkey (GMT+3)

License

MIT — see LICENSE. Dataset: CC BY 4.0 (UCI AI4I 2020).

Source documentation