Model card
About, in the format he already writes.
He wrote a model card for the segmentation model that shipped — intended use, evaluation, and a limitations section nobody asked for. This page is the same document, about the engineer instead of the model.
Model details
- Name
- Muhammad Mughees Ul Haq
- Version
- Two years in production, June 2024 → present
- Architecture
- Computer engineer by training, machine learning engineer by practice. Embedded systems underneath, computer vision on top.
- Location
- Lahore, Pakistan. Working remote with teams in Europe.
- Education
- BS Computer Engineering, Information Technology University of the Punjab, 2021–2025.
- Certifications
- AWS Cloud Foundations · AWS Data Engineering · Fast.ai Deep Learning · CS50 Python
Intended use
- Primary
- Owning a vision or ML pipeline end to end — dataset design, training, an evaluation harness that can actually reject a model, ONNX export, and the service around it.
- Also effective
- Turning a manual or no-code process into a retry-safe, idempotent system. Building the harness that decides which of four candidate pipelines ships.
- Out of scope
- Front-end product design, or research aimed at publication rather than deployment.
Training data
- Production
- Face-blur segmentation, monocular-depth fill-level estimation, keypoint localization, open-vocabulary detection, OCR and barcode reading, VLM-assisted annotation at volume.
- Infrastructure
- Docker, AWS EC2 and S3, systemd, ONNX Runtime, FastAPI. Long training jobs run under tmux with cron auto-resume, because the power goes out.
- Cross-site
- Ran QA cycles and cleared deployment blockers with a European partner across time zones.
- Before that
- FreeRTOS task design on ESP32, OBD-II telemetry, host-side C unit tests, hardware-in-the-loop validation on a physical rig.
Evaluation
- Method
- Held-out test sets, benchmarked against a named baseline, with the configuration recorded alongside every number.
- Frame segmentation
- F1 49.8 → 84.7 across eight generations. Shipped model 81.2 at imgsz 1792, conf 0.35, on a 225-image internal test set.
- Fill level
- MAE 18.57 → 2.14 on a labelled holdout; R² 0.9916 in validation.
- Keypoints
- COCO AP 0.9804 verified; validation AP 1.000 at epoch 70.
- Vehicle health
- Damage mAP@50 84.3%, parts mAP@50 87.6%.
- Agent
- 100% on a 10-query routing benchmark with a hand-specified expected tool per query. Sub-2s direct responses.
Limitations
- Scale
- Production experience is two years. Pipelines have been evaluated on hundreds to thousands of images, not at very high request volume.
- Benchmarks
- Most evaluation is on internal test sets built for the task, not on public leaderboards. The numbers here are real, and they are not directly comparable to published benchmark figures.
- LLMs
- Foundation models are integrated, prompted, routed and served — not trained or fine-tuned from scratch.
- MLOps depth
- Comfortable with Docker, ONNX, EC2 and evaluation harnesses. Has not operated a large managed platform with feature stores and a formal model registry.
- Ownership
- Has owned pipelines and their evaluation. Has not led a multi-team engineering organisation.
Ethical considerations
- Privacy
- The flagship project exists to blur faces, not to find them. Its model card says so explicitly: it should not be used to locate faces for identification.
- This site
- Three of six annotation projects are held back from publication because they contain identifiable faces from client CCTV and photography. Their counts appear in the totals; their images do not.
- Attribution
- CarDD (Wang et al., IEEE T-ITS 2023) is used for non-commercial research with credit. carparts-seg (Ultralytics) is AGPL-3.0.
- Clients
- Client names are anonymized throughout. Every metric is unchanged.
Contact
Open to machine learning and computer vision roles.