Undergraduate researcher in trustworthy machine learning, computer vision and edge AI. BS Computer Science, FAST-NUCES Islamabad.
I work on the problem of models that are confidently wrong, and on the systems that have to survive that: uncertainty quantification, calibration, out-of-distribution behaviour, and human-in-the-loop escalation when a prediction should not be trusted.
Risk-Aware Active Learning for Skin Lesion Classification · manuscript under peer review, 2026 L. Noor, M. D. Asif, H. Ramzan, A. Atiq, M. A. Saeed, A. Jamil, A. Din
A dual-metric human-in-the-loop escalation policy for dermoscopy: a case is routed to a clinician when either model uncertainty or a dedicated clinical-risk head crosses a per-round calibrated threshold. This targets the failure standard active learning cannot see, where a model is confidently wrong about a melanoma and auto-accepts it. We prove the escalation set is a superset of the uncertainty-only set, so unsafe auto-acceptance cannot increase.
Across 24 experiments on HAM10000: 43% fewer unsafe auto-accepts in 12 of 12 matched
configurations, at 9.1% more oracle labels (Holm-corrected p = 0.003), with no significant change in F1-macro
(p = 0.305). Validated out-of-distribution on 14,885 ISIC-2019 images with HAM10000
duplicates removed: 83.3% accuracy win rate, 75.0% melanoma-safety win rate.
→ RiskAware-ActiveLearning
Hallucination-Free Retrieval-Augmented Generation · undergraduate research assistant, FAST-NUCES
An agentic RAG question-answering system for Classical Arabic: GATE Arabic embeddings fine-tuned with Matryoshka loss, an HNSW index over 12,472 provenance-tagged passages, and a bounded retrieve → sufficiency-check → rewrite loop with per-step citation validation. I build the evaluation framework, including a knowledge-graph guardrail that extracts subject-relation-object triples from source and answer to reject unsupported claims.
| ISBGlyph | Smart-city simulation of Islamabad in C++/SFML with every data structure written from scratch — adjacency-list graphs with Dijkstra routing, n-ary trees, heaps, open-addressing hash tables. Zero STL. |
| PBSVertex | Pakistan's Consumer Price Index across 17 cities modelled as a price-similarity graph, using cosine similarity and centrality analysis to identify economic hubs. Live |
| Super Mario (Inspired) | A playable four-level platformer in pure x86 Assembly: hand-rolled memory management, stack-based collision detection, custom real-time sound manager. No engine, no standard library. |
Production work at DenseFusion (geospatial ML platforms, YOLO11n on NVIDIA Jetson, offline-first agricultural advisory systems) is under NDA. One deployment is publicly viewable at serena.com.pk/green.
Languages Python · C++ · Java · JavaScript/TypeScript · SQL · x86 Assembly ML PyTorch · scikit-learn · Ultralytics YOLO · active learning · uncertainty quantification · calibration · OOD detection · Grad-CAM++ / Score-CAM · RAG · HNSW retrieval Vision & geospatial Rasterio · GDAL · PROJ · QGIS · GeoPandas · Cloud-Optimized GeoTIFF · PostGIS · TiTiler Edge NVIDIA Jetson Nano / Orin Nano · MAVLink telemetry · on-board inference under power and thermal constraints Systems FastAPI · Next.js · NestJS · PostgreSQL · Redis · RabbitMQ · MinIO · Docker · Nginx · Linux
mdyenasif@gmail.com · LinkedIn · Islamabad, Pakistan