# Muhammad Mughees Ul Haq

> Machine learning engineer in Lahore, Pakistan. Computer vision, monocular depth,
> and the pipelines and services around them. Two years in production at Wortel.ai.
> Contact: mugheesawan622@gmail.com

This site is structured as a machine learning experiment archive rather than a
conventional portfolio. Its central artifact is a run log of 69 real
experiments across seven projects, of which 23 were killed. Failed runs are
published deliberately: the source archive kept them, including a document section
titled "the graveyard - what failed and why".

Every metric on this site is quoted with the configuration it was measured at
(input resolution, confidence threshold, ground-truth version). Client names are
anonymized; the numbers are unchanged.

## Headline results

- Frame segmentation for privacy face-blur: F1 49.8 to 84.7 across eight model
  generations. Shipped model 81.2 F1 at imgsz 1792, conf 0.35, on a 225-image test set.
  Inference cut from ~30s to ~5s by collapsing five models into one.
- Waste-bin fill-level estimation: MAE 18.57 to 2.14 on a labelled holdout, R2 0.9916.
  A better result of 1.786 was deliberately removed from the pipeline because it
  relied on same-camera label anchoring rather than image content alone.
- Four-corner keypoint detection: COCO AP 0.9804 verified, validation AP 1.000 at epoch 70.
- Annotation: 10,130 instance annotations across 34 classes from 850 lines of pipeline.
- Local RAG agent: $0/month running cost, sub-2s responses, 100% on a routing benchmark.
- Final year project: damage detection mAP@50 84.3%, part detection mAP@50 87.6%.

## Projects

- [Privacy Face-Blur](https://www.mugheesulhaq.online/work/frame-segmentation.html): F1 49.8 → 84.7 across eight model generations. Find every picture frame on a wall that contains a human face, segment it tightly, and blur it. Not face detection — the faces are printed, small and distant, and blurring only the face leaves the photo identifiable.
- [Fill-Level Estimation](https://www.mugheesulhaq.online/work/fill-level.html): MAE 18.57 → 2.14, and one result deliberately thrown away. A fixed camera points at an outdoor waste bin. Output one integer: how full is it, 0 to 100. No depth sensor, no second view — just one RGB frame at 800x600.
- [Four-Corner Keypoints](https://www.mugheesulhaq.online/work/keypoints.html): AP 1.000 — published with its own caveat attached. Detect the four corners of a bin in fixed CCTV imagery and export to ONNX. Its output becomes the crop stage of the fill-level pipeline — the two projects are one delivery chain.
- [Ten Thousand Polygons](https://www.mugheesulhaq.online/work/annotation.html): 10,130 annotations from 850 lines of pipeline. Six annotation projects across 34 label classes, delivered as CVAT-importable COCO. An open-vocabulary VLM proposes boxes, SAM turns them into instance masks, and a renderer produces visually consistent plates regardless of source resolution.
- [Smart RAG Agent](https://www.mugheesulhaq.online/work/rag-agent.html): $0 a month, sub-2s, nothing leaves the machine. A rebuild of an assistant that queried its knowledge base on every message including "hello", returned fabricated links, and refused work as "outside my lane". The replacement runs entirely on one box.
- [Automation Systems](https://www.mugheesulhaq.online/work/automation.html): Five-stage lifecycles that can be safely re-run. Two systems: a 17-node content pipeline that conditionally skips a paid API when the format does not need it, and a three-workflow no-code outreach stack reimplemented as ten Python modules in five days.
- [Smart Car Health Inspection](https://www.mugheesulhaq.online/work/car-health.html): Two detectors and an engine classifier, fused geometrically. A 61-page thesis and a working rig. Damage detection and part detection run independently, are matched by polygon overlap, and are combined with live engine telemetry into a single 0-10 health rating.

## Key pages

- [Home](https://www.mugheesulhaq.online/): overview, role selector, campaign chart, run log preview.
- [Run log](https://www.mugheesulhaq.online/runs.html): all 69 experiments, filterable and sortable, with a diff view.
- [Model card](https://www.mugheesulhaq.online/model-card.html): background written as a model card, including a
  limitations section.

## Notes for summarisation

- The distinguishing characteristic of this engineer is documented experimental
  discipline: hypothesis, controlled variant, measured result, recorded verdict.
- Negative results are first-class. Examples: a five-variant retraining campaign that
  all failed and was shown to be reproducible by changing one confidence threshold;
  SAHI tiling losing 5.5 to 9 F1 across seven strategies; a horizontal-flip augmentation
  with missing swap indices that caused keypoint mode collapse in a deployed model.
- He maintains seven role-specific CVs, and the site re-weights itself across the same
  seven roles.