I will build custom computer vision models


About this gig
I train computer vision models that work on your images, not just on benchmark datasets.
What I build:
- Image classification (defects, quality, products, plants)
- Object detection and counting (safety gear, retail, agriculture)
- Segmentation and OCR-style pipelines
- Models compressed to run on phones, Raspberry Pi and edge devices
Proof of depth: I compressed a 50M-parameter vision model to 11M parameters while keeping 98% of peak AUC (case study on theabm.dev).
You receive: Trained weights and clean PyTorch code · Evaluation report (precision, recall, mAP/AUC, confusion matrix, failure examples) · Inference script or small API + README.
Licensing: I default to permissively licensed models (RF-DETR, YOLOX, torchvision) so you can use the result commercially. Ultralytics YOLO is AGPL-3.0 I use it only if you hold its license or accept AGPL.
Honest limits: I train on a 16 GB GPU, so I fine-tune efficient architectures rather than pretrain huge ones.
Message me with your goal, sample images and how many labeled images you have.
Get to know Afnan Mustafij
Machine Learning Engineer
- FromBangladesh
- Member sinceJul 2026
- Avg. response time1 hour
Languages
English, Bengali
FAQ
I have no labeled data. Can you still help?
Yes, I can guide labeling with Roboflow/CVAT or use public datasets where legally suitable. Annotation is quoted separately.
Can you guarantee a specific accuracy?
No honest engineer can before seeing the data. I run a quick feasibility check first and tell you realistic targets.
Who owns the model and code?
You do, with full commercial rights after final payment.
