I build benchmarks, protocols and audits that test whether a model's reported performance survives deployment, anchored in computer vision and extended to personalization. In vision, that means longitudinal benchmarks whose protocols separate identity leakage from genuine distribution shift; in personalization, audits of whether a recommender's explanation matches what it actually did.
It started with a cow. In 2019 I built a model that estimates a cow's weight and breed from photographs, and it was second runner-up for AI in Agriculture at Bangladesh's BASIS National ICT Awards. What nobody asked was whether it would still work a year later, on animals it had never seen. I spent the next six years collecting the data to find out.
By day I'm a Staff ML Engineer at Graaho Technologies in Dhaka, where I lead the machine learning and backend team. Almost every question I research began as a problem in production.
Research
- BoviShift (first author, project lead): a six-year benchmark of 2,657 animals that splits an accuracy drop into identity leakage and temporal shift. Code and data will be released publicly. Two follow-up studies with the same team are under way.
- Auditing LLM-based recommenders (in preparation): does the stated explanation match the ranking, does the model invent item attributes, and do those failures fall unevenly across users? Runs without GPU compute.
- CID (ICCA 2022): 513 cattle across 8 breeds, 2,052 photographs and 15,812 video frames, released with regression and classification baselines. Paper
In production, one rule each
| System | The rule it enforces |
|---|---|
| AlgoRec, a recommendation platform on AWS Marketplace | No model goes live until it beats a popularity baseline on AUC |
| Koronik, customer-service agents for Islami Bank Bangladesh, 5,000+ people a day | Below a confidence threshold, the agent stops and hands the customer to a person |
| Margin, an adaptive writing tutor I built for myself | An LLM grades; deterministic code decides which scores the learner may see |
| Easymator, construction takeoff from drawings | Wall detection stays manual, because its mistakes would be invisible to the contractor |
Each rule is a small evaluation instrument inside a product. My research asks how to make them general. Margin's source is private; access on request.
Open source worth a look
- claude-standing-orders: an orchestration template for agentic coding in which authorization boundaries fail closed.
- traffic-ai: vehicle analysis and toll collection from live video (demo).
- Cow-weight-and-Breed-Prediction: the code behind CID and the original Smart Guess models.
- Categorized-Affect-Map: my undergraduate thesis, published at ICCA 2020.
Also
I translated Week 12 of the NYU Deep Learning course (Yann LeCun and Alfredo Canziani) into Bengali. AWS Certified Machine Learning, Specialty. Kaggle 3× Expert.
Most of my day-to-day work is in PyTorch, FastAPI, MLflow, Airflow and AWS (SageMaker, Bedrock).
If you work on evaluating vision systems, recommenders or agents, in a lab or on a product team, I'd like to hear what you're measuring.
Find me: mobasshirbhuiya.com · Google Scholar · ORCID · Semantic Scholar · DBLP · LinkedIn · Kaggle · mobasshir.bhuiya@gmail.com