Md. Asif Uddin
Deep learning researcher — vision, medical imaging, vision-language models, causal inference, bioinformatics and gene sequence analysis
Research and publications
HierarchiRetina
2026HierarchiRetina: A Diabetic Retinopathy Screening and Grading Cascade with Heterogeneous Mixture-of-Experts Lesion Segmentation
under review Co-authors: Niloy Farhan, Rafeed Rahman, Md. Golam Rabiul Alam.
IEEE JBHI · Preprint on arXiv
End-to-end quadratic weighted kappa 0.804 · Referral sensitivity 0.832 · Stage I gate, AUC 0.929 · Lesion Dice, pooled 0.184 – 0.579 · Vessel Dice 0.731 · Official DDR test set, six classes, end to end 85.6%
Rank Radii Transfer as Quantiles: Taxonomic Placement and Sub-Taxonomy Construction for Unseen Species
under review Supervisor: Niloy Farhan.
Beyond Multi-Stage Pipelines: A Label-Correlation Ensemble for Multi-Label Retinal Fundus Disease Classification
preprint Co-authors: Jannatul Ferdous, Asif Islam, Zobaer Mahmud Zisan, Promit Ghosh, Shah Ahafik Arman, Niloy Farhan, Rafeed Rahman.
arXiv · cs.CV · Preprint on arXiv
Exact match 0.694 · Micro-F1 / macro-F1 0.678 / 0.456 · Macro-AUC 0.972 · Detection rate 0.689 · Disease lost at the cascade's two gates 34.2%
Measure First, Write Second: Verified Generation of the Kidney Section of a Radiology Report from Abdominal CT
in preparation Supervisor: Niloy Farhan.
HCGT-PG
2026HCGT-PG: Hierarchical Counterfactual Graph Transformer, Perturbation-Grounded — causal biomarker discovery in Alzheimer's disease microglia
proposal accepted Supervisor: Dr. Badhan Das.
Software
ZenodoMd Asif Uddin. doi: 10.5281/zenodo.23741643 · repository
Education
MSc, Computer Science and Engineering
Fall 2026 —BRAC University, Dhaka. Thesis proposal accepted: HCGT-PG. Supervisor: Dr. Badhan Das.
BSc, Computer Science and Engineering
Fall 2021 – Summer 2026BRAC University, Dhaka. CGPA 3.31 / 4.00. Thesis: HierarchiRetina, supervised by Rafeed Rahman; not publicly released; the work is now under review at IEEE JBHI, with a preprint on arXiv. Coursework in machine learning, deep learning, compilers, theory of computation, DBMS, operating systems, algorithms and software engineering.
Higher Secondary Certificate
2020Uttara High School and College, Dhaka. GPA 5.00 / 5.00. Science.
Secondary School Certificate
2018Ullapara Merchant Pilot High School, Sirajganj. GPA 4.72 / 5.00. Science.
Technical
- Languages
- Python (primary). TypeScript and JavaScript (Next.js, React, React Native). Rust (Tauri). SQL (PostgreSQL). C (coursework level).
- Deep learning
- PyTorch, timm, torchvision, Hugging Face Transformers, PEFT, segmentation_models_pytorch. Training: mixed precision (bf16/fp16) with GradScaler tuning, AdamW with warm-up and cosine decay, gradient accumulation and clipping, EMA, early stopping on a validation monitor, k-fold CV with resumable per-fold checkpointing. Fine-tuning: LoRA and 4-bit QLoRA on language and vision-language models, frozen-backbone transfer, linear probing against partial and full fine-tuning. Imbalance and calibration: pos_weight, weighted samplers, source-balanced negative sampling, per-class and out-of-fold thresholds tuned on validation only. Inference: test-time augmentation with re-tuned thresholds, probability ensembling with weights chosen on validation. Building blocks: ordinal regression (CORN), mixture-of-experts, cross-attention fusion, attention gates, ASPP, GeM pooling, deep supervision, Grad-CAM. Numerical checks: layer-by-layer equivalence against reference weights, and finding a bf16 softmax that collapsed scores.
- Architectures used hands-on
- Vision: ConvNeXt V2, Swin Transformer V2, Swin UNETR V2, DINOv2/v3, EfficientNet, and a Vision Transformer built from scratch and verified layer by layer against ViT-Base. Segmentation: U-Net, U-Net++, Attention U-Net, SwinHRUNetPP, HSMoE-AUNet (sparse mixture of heterogeneous experts). Heads and fusion: transformer decoders with learned label queries, gated cross-attention between masks and images, frozen image-text encoders with a trained fusion module. Language and vision-language models: Qwen3-VL-2B fine-tuned with QLoRA, Qwen2.5-7B with QLoRA and Qwen2.5-3B with LoRA; gpt-oss-120b, Qwen2.5 Coder 32B and Qwen2.5 3B served in deployed systems. Retrieval and speech: BGE embeddings, MiniLM cross-encoder reranking, multilingual MiniLM embeddings, Whisper large-v3, Kokoro-82M.
- LLMs and retrieval
- Serving and inference: vLLM with prefix caching, Ollama, ONNX Runtime, Groq, Cloudflare Workers AI. Retrieval-augmented generation: structure-preserving PDF parsing (pdfplumber), hybrid BM25 and dense retrieval, reciprocal rank fusion, cross-encoder reranking, multilingual embeddings, exact NumPy vector search, incremental re-indexing. Grounding: citations computed from the retrieved passages, quotations verified against the source before display, refusal when the evidence is not there. Speech and OCR in the pipeline: Whisper, Kokoro, Tesseract.
- Medical imaging
- Fundus photography, non-contrast CT, NIfTI and DICOM handling, FOV masking with morphological erosion, CLAHE on the green channel, retinal cropping, patch extraction and stitching, lesion-level connected-component evaluation, de-identification, radiologist report parsing.
- Evaluation
- Classification and grading: QWK, ICC, AUC-ROC, AUC-PR, sensitivity/specificity trade-offs, referral sensitivity. Multi-label: exact match, Jaccard, micro- and macro-F1, macro-AP, detection rate. Segmentation: Dice, IoU, HD95, lesion-level and image-level recall. Clustering: NMI. Retrieval and RAG: Recall@k, MRR, hand-labelled ground truth, verified-quotation rate, refusal on unanswerable questions. Screening: WSS@95%, recall@10%. Statistics: confidence intervals, paired bootstrap with Holm correction, repeated paired seeds. Method: end-to-end evaluation of staged systems with per-stage error attribution, ablations including mask ablation, within-source metrics and source-balanced negatives against dataset shortcuts, baselines computed rather than cited, held-out test discipline and leakage assertions.
- Causal inference and computational biology
- Causal graphs and DAGs, d-separation, interventions, counterfactual reasoning, gene perturbation analysis, gene regulatory network inference, multi-omics integration and biomarker discovery.
- Data
- pandas, NumPy, scikit-learn, OpenCV, Pillow, Albumentations, Matplotlib, Seaborn, nbformat. Merging heterogeneous public datasets into one label space while keeping each source's provenance and official split; datasets worked with include EyePACS, APTOS 2019, DDR, IDRiD, Messidor-2, FGADR, RFMiD 1.0 and 2.0, ODIR-5K, iNaturalist, MedMNIST, CIFAR-10 and SYNERGY. Synthetic benchmark generation (2,800 diagrams with 25,698 structural questions); PDF and statute text parsing; citation formats (RIS, MEDLINE, BibTeX) and deduplication.
- Backend and deployment
- FastAPI, REST APIs, SQLAlchemy and Alembic. PostgreSQL with row-level security and full-text search, Prisma, Redis. Authentication with JWT, OIDC sign-in, two-factor sign-in and passkeys. Next.js and React front ends. Docker and Caddy; AWS S3 and CloudFront, Cloudflare Workers and R2, Vercel, Hugging Face Spaces. Testing: pytest, Vitest, Playwright end to end, mypy --strict, CI on every push.
- Model familiarity
- Beyond the models I have used hands-on, I study current models closely enough to write a note on each: architecture, training objective, scaling behaviour, benchmarks, limitations and practical use. Vision: ViT, ConvNeXt V2, EfficientNet V1 and V2, DINOv2 and DINOv3 (including registers), SAM 3. Multimodal: SigLIP 2, EVA-CLIP 8B and 18B. Medical imaging: nnU-Net, Swin UNETR V2, MedGemma. Language: the Qwen family including Qwen3.8, Llama 3.3 70B and Llama 4, DeepSeek V4, Kimi K3 and K2.6, and sparse mixture-of-experts design. Biology and chemistry: AlphaGenome, Evo 2, Boltz-2, Geneformer, scVI and the STATE virtual-cell model. I do not count familiarity as hands-on experience; the notes are published on asifuddin.com.
- Other
- Git and GitHub Actions, Jupyter and Colab, LaTeX, lex/yacc, WebAssembly toolchains (Pyodide, clang), CUDA-level debugging.
Languages. Bangla, native. English, full professional proficiency and the medium of instruction throughout undergraduate study.
AI systems
ResearchLens — evidence-grounded literature retrieval
Evidence-grounded RAG over 106 papers, 15,664 passages, plus a live-fetched index of my own site and textbook (Python, FastAPI, ONNX, Gradio). Structure-preserving PDF parsing; hybrid BM25 + dense retrieval with RRF fusion and cross-encoder reranking; mechanically enforced grounding — computed citations, fabricated-marker removal, refusal when unsupported. Per-session document upload isolated from the shared index; subset-scoped retrieval; related-work discovery across arXiv, PubMed and OpenAlex; benchmark harness with hand-labelled ground truth. 293 tests, deployed on Hugging Face, runs locally with no API key.
LocalScholar — a private, laptop-scale research assistant
Local-first RAG research assistant (FastAPI, React, ONNX, Ollama). Hybrid BM25 + dense retrieval with RRF fusion and cross-encoder reranking; measured ablation to Recall@5 0.933, MRR 0.840; verified refusal on unanswerable questions; citations computed from passage overlap rather than self-reported. 100 tests, runs offline on 8 GB.
Criminal Law Q&A Assistant — answers on Bangladesh criminal law, grounded in the statute and checked against it
Source-grounded legal Q&A over Bangladesh's Code of Criminal Procedure, its Schedule II and the Penal Code (FastAPI, Next.js, ONNX, Groq gpt-oss-120b with an Ollama fallback, whisper-large-v3, tesseract). Verbatim excerpts verified against the statute before display, Schedule II lookup for offence classification, amendment history on every citation, and incremental re-indexing. Text, speech, image and document input in English and Bangla. Evaluated in four stages on 101 questions: verified quotations rose from none without retrieval to 87.5%. Live on Hugging Face.
PRISMA-Local — screening a systematic review on your own hardware
Local LLM screening assistant for systematic reviews (PyTorch, PEFT, vLLM, Gradio). QLoRA 4-bit fine-tune of Qwen2.5-7B and LoRA fine-tunes of Qwen2.5-3B scoring title+abstract against a protocol in one token, ranked by logit margin; trained on 67 SYNERGY reviews, evaluated on 6 held out at natural prevalence (13,665 records, 4.6% positive). Cross-topic transfer with zero in-review labels: WSS@95% 0.331 vs 0.129 zero-shot, reported level with a TF-IDF baseline and below active learning, both computed rather than cited. Diagnosed a bf16 softmax collapsing 13,665 records onto 45 distinct scores — recall@10% +26% after an fp32 fix that moved AUC by 0.002 — and made score resolution a logged, failing condition. vLLM prefix caching 1.40x. In-browser demo, RIS in and RIS out.
Vi-Graph — diagram images turned into editable, queryable graphs
Diagram-to-graph reconstruction with a fine-tuned vision-language model (PyTorch, Transformers, PEFT/QLoRA, FastAPI, Next.js, React Flow). Qwen3-VL-2B with 4-bit QLoRA: 17.4M trainable parameters, 1.4 hours, 7.6 GB peak memory. Built a 2,800-diagram synthetic benchmark with unseen test layouts and 25,698 structural questions. On 500 held-out diagrams: graph similarity 0.581 → 0.740, edge F1 from 0.499 to 0.638, structural QA 0.437 → 0.582, valid graphs 72% → 91%, all p < 0.001 (paired bootstrap, Holm). Reported an OCR + OpenCV baseline that ties on graph similarity (0.738); diagnosed runaway enumeration on dense diagrams and added a guard that cut evaluation time 26% with no significant score change. Schema-validated output with logged repair, graph-grounded QA, exports; 767 tests.
Lucy — an AI desktop companion that runs entirely on the Mac
macOS AI desktop companion (Tauri 2, Rust, TypeScript, three.js/VRM, Apple Speech, Ollama, Kokoro TTS on ONNX). Code-driven 3D character with spring-bone physics and six moods; hands-free wake word on Apple's on-device recogniser; local neural voice whose expression tags drive face and pacing. Rust owns all native access and every network call, keys in the Keychain, and the webview's CSP permits no network; any OpenAI-compatible provider or Anthropic, local and free by default. Click-through transparent overlay via a hit-test thread, cursor petting from a Rust watcher, a deterministic calculator in place of model arithmetic, gapless speech by synthesising ahead. 12–30 fps idle, ~190 MB voice server faster than real time. 146 Vitest and 48 Rust tests, clippy -D warnings, CI on Linux and macOS.
Officina AI — an execution tracer for Python, C and Java, with a tutor that explains the trace
Execution tracer for Python, C and Java that runs in the browser (Pyodide, tree-sitter, clang in WebAssembly, the Eclipse compiler on CheerpJ, Cloudflare Workers AI). Python traced under sys.settrace with AST-instrumented conditions; C and Java instrumented at the source with insert-only edits that keep every line number, compiled in the browser and traced by their own runtimes — C variables read back by address, pointers named by their targets, out-of-bounds, divide-by-zero and NULL caught on the line. One step format for all three, checkpointed so any of 50,000 steps shows in under 1 ms. LLM tutor grounded in the recorded facts, behind a same-origin, rate-limited Worker. Every C and Java example traced natively in CI against its untraced output.
AI projects
Glass-Box ViT — a Vision Transformer rebuilt from scratch and verified layer by layer against a pretrained ViT-Base
From-scratch Vision Transformer in PyTorch, verified against a pretrained reference (PyTorch, timm, torchvision, scikit-learn, MedMNIST, Colab T4). Loads all 152 tensors (86,567,656 parameters) of a pinned, SHA-256-verified timm ViT-Base and matches it exactly at all 100 intermediate stages (max difference 0.0) on an M1 CPU and a Tesla T4; 81.0 ms on CPU and 15.0 ms on T4 per 224-px image, against 14.6 ms for timm's fused attention. Frozen-protocol CIFAR-10 ablations over three paired seeds: patch size matters most (75.3% to 59.3% validation accuracy from patch 4 to 16 at width 192); learned positions and mean pooling win in every seed; head count makes no clear difference. RetinaMNIST transfer, a research exercise: test QWK 0.750 linear probe, 0.775 ± 0.038 partial and 0.773 ± 0.038 full fine-tuning. 34 tests.
MONAI — the metric that hid the failures
Open-source contribution to MONAI (Project-MONAI/MONAI, 8.7k stars), the standard medical imaging framework. Found a silent correctness bug in HausdorffDistanceMetric: a prediction that misses the structure entirely returns nan at every percentile where the maximum returns inf, and nan is excluded from the reduction, so failed cases are dropped from a dataset HD95 rather than scored. Demonstrated that the reported metric does not move as a model degrades from 0 to 99 misses in 100 images. Traced to torch.quantile interpolating between two infinities; fixed with a guard and twelve regression tests, verified failing before and passing after. Reported as issue #9095; the fix, PR #9096, was reviewed and merged into MONAI on 6 October 2026, closing the issue.
Food Expiry Prediction — CNN and Random Forest models that estimate when food will spoil
Food expiry prediction to cut food waste. A CNN judges freshness from photographs, and Random Forest and MLP models weigh temperature, humidity and storage duration. Built on the Kaggle Fruits Fresh and Rotten dataset with engineered features such as days since purchase.
Other projects
Winnow — a free, self-hostable platform for systematic reviews
Free, open-source, self-hostable platform for systematic, scoping and rapid reviews (FastAPI, PostgreSQL, Redis, React, TypeScript, Docker, Caddy). Import from six citation formats, automatic deduplication, dual blind screening with conflict resolution, relevance ranking that retrains per reviewer with a stopping estimate, open-access full-text retrieval with virus scanning, dual data extraction with consensus, four risk-of-bias instruments, and PRISMA 2020 output with Cohen's and Fleiss' kappa. Authorisation checked per review with PostgreSQL row-level security behind it and blind mode enforced server-side; Argon2id, 2FA, OIDC sign-in, append-only audit log, OWASP ZAP baseline clean in CI. Measured against written budgets on 100,000-record reviews: screening p95 32–69 ms, import 49 s, deduplication of 50,000 in 13 s, LCP 0.7 s on 4G; 50 concurrent reviewers at p95 37 ms with 0 errors in 19,920 requests. WCAG 2.2 AA with no serious axe findings across 43 pages; 900 backend tests at 95% coverage, Playwright e2e, mypy --strict. Specified as a ten-phase plan with per-phase acceptance criteria and built with AI coding agents against those gates; ~150 commits.
SANDHI Research Lab — a public website, member workspace and administration in one application
Research lab platform, live at sandhiresearch.org (Next.js 16, React 19, TypeScript, PostgreSQL/Prisma, Better Auth, Cloudflare R2, Resend, Vercel). Public site with a research map, project progress, publications and Bengali-aware full-text search; member workspace with task boards, meeting notes, an append-only experiment log and Markdown drafting with review; sixteen administrative managers with an audit log. Invitation-only accounts, mandatory 2FA with passkeys, breach-checked passwords, audit written in the same transaction as each change, signed links for private files. Data-layer caching invalidated by every edit under a nonce-based CSP; search on PostgreSQL full-text indexes. 101 pages, 48 data models, 400+ unit and 140+ Playwright tests including security boundaries and accessibility; first commit to production in 14 days.
SANDHI app — the research lab's Android app, a second client of the website rather than a second system
Android app for SANDHI Research Lab, released as a signed APK (Expo SDK 57, React Native 0.86, React 19, TypeScript, Expo Router, expo-secure-store, EAS Build). Public research map, projects, papers, people, news and search; member sign-in with password and two-factor code, the token in the device keychain and sign-out ended on the server. No backend of its own: it reads the website's /api/v1 under the same visibility rules as the site's pages. Minimum-version gate set from the site, published SHA-256, and an ARM-only build that halved the APK from 101 MB to 57 MB, asking Android for network access only. iOS built and tested, awaiting App Store funding.
Officina — a notebook that compiles and runs Python, C, C++ and Java in the browser
Browser-only notebook for Python, C, C++ and Java (Pyodide, clang and wasm-ld compiled to WebAssembly, the Eclipse compiler on CheerpJ). Python runs in a Web Worker with the network APIs removed, with Stop, a 30 s limit and a 256 KB output cap; C and C++ are genuinely compiled and run in the browser; notebooks persist locally and export as .ipynb. No server; the example library is edited from the site's CMS.
CSMMS — Campus Student Management and Marketplace System
FastAPI backend: REST API design, filtered and paginated marketplace, S3 and CloudFront image upload, service impact metrics, SQLAlchemy, JWT, Alembic.
Compiler symbol table — scoped name lookup for a lex and yacc/bison front end
Scoped symbol table for a compiler front end using lex and yacc/bison, with cross-platform builds on Linux and macOS.