asifuddin.com · Papers · CV · LinkedIn · ORCID
Medical imaging, vision-language models and causal inference. Every number below was measured, and the repositories show how.
| Work | Status | |
|---|---|---|
| I | HierarchiRetina: lesion-aware diabetic retinopathy grading, 0.9422 macro-AUC across six public datasets | BSc thesis, deposited |
| II | Rank Radii Transfer as Quantiles: placing unseen species inside the taxonomy | under review |
| III | Measure First, Write Second: CT kidney reports that cannot state an unmeasured value | in preparation |
| IV | HCGT-PG: causal biomarker discovery in Alzheimer's microglia | MSc proposal accepted |
- ResearchLens: evidence-grounded retrieval over 103 papers. Every claim cites a passage, and it refuses when the evidence is missing. 293 tests. Live demo
- LocalScholar: a private paper assistant that runs offline on 8 GB. Recall@5 0.933, MRR 0.840. Recorded session
- PRISMA-Local: QLoRA/LoRA Qwen screeners for systematic reviews. Found that bf16 collapsed 13,665 scores onto 45 values. The fix raised recall@10% by 26% while AUC moved by only 0.002. Demo
- Winnow: a self-hostable platform for systematic reviews, from import to the PRISMA 2020 report.
In MONAI, HausdorffDistanceMetric(percentile=95) returned nan for a missed prediction, so the average silently dropped those images. A model that missed the lesion in 99 of 100 images reported the same HD95 as one that found it every time.
Reported in #9095. The fix is under review in #9096, and the full analysis is in A-PR.
PyTorch timm MONAI PEFT vLLM ONNX FastAPI Docker TypeScript