Md. Asif Uddin

Md. Asif Uddin

AI and machine learning engineer — I build systems that train, run and can be checked.

Correlation is cheap. Only an intervention costs anything. 

Dhaka, Bangladeshgithub.com/asifuddin01CV

Figura IThree runs, one surfaceTwo start together and differ only in momentum. They do not finish in the same basin, and nothing inside a run tells it that a better minimum lies three tenths away.plain descentwith momentumlarger step

Now · September 2026

Four things running. Two build something; two work out what the literature already settled, so the first two are not arguing with a field they have not read.

  1. HCGT-PGCausal biomarker discovery in Alzheimer's microglia — what happens to a cell when a gene is knocked down, checked against CRISPR screens. Proposal accepted; MSc begins this term.
  2. Renal reportingMeasure first, write second — the kidney section of a CT report, generated from computed values a verifier will not let the model drop.
  3. Review — validationHow perturbation-prediction methods are validated, and against what. Protocol stage, PRISMA-ScR.
  4. Review — renal CTWhich structures and tasks the CT literature automates, and where it stops. Protocol stage, five databases.

Systems that run

Not repositories — things a stranger can open and put a question to. Both ground every claim in a passage you can check, and both refuse when the evidence is not there, which is the part that takes the work.

  • ResearchLens — evidence-grounded literature retrieval

    Evidence-grounded RAG over 101 papers, 9,593 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. 261 tests, deployed on Hugging Face, runs locally with no API key.

    Open ResearchLens

  • 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.

    LocalScholar on GitHub

Architectures built, not fine-tuned

Each of these exists because an off-the-shelf model was wrong about the problem in a way that could be named. The rejected list matters as much as the kept one — it is the difference between a choice and a default.

  1. HSMoE-AUNet

    A microaneurysm is a few dark pixels. A cotton wool spot is a soft pale patch. They do not respond to the same design.

    Hierarchical sparse mixture-of-experts attention U-Net on a ConvNeXt-Small encoder, with a different expert set per lesion type. Custom experts: MultiScaleExpert, BrightSpotExpert, EdgeEnhancingExpert, SoftRegionExpert, TextureExpert.

  2. SwinHRUNetPP

    Vessel segmentation needs global context and hairline spatial precision at once, and most hybrids trade one away.

    A Swin-based high-resolution U-Net++ hybrid, arrived at by rejection rather than by design.

    Rejected —SwinV2-Large, smp UNet++, plain SwinUNet, Lovász loss, BoundaryLoss, warm-restart scheduling, label smoothing, encoder freezing

  3. MultiScaleMANet

    Microaneurysms vanish at low resolution and lose context at high resolution.

    Three branches — full 512×512 through EfficientNet-B4, four 256×256 patches through a shared EfficientNet-B0, sixteen 128×128 patches through the same encoder — with spatial reassembly and bottleneck fusion. Alignment was proved on paper before implementation.

  4. RichAttentionUNet

    A baseline worth trusting has to be one you can account for line by line.

    Written from scratch with no pretrained weights: residual conv blocks with SE channel attention, an ASPP multi-scale bottleneck, attention gates in the decoder, deep supervision heads.

  5. Custom losses

    Off-the-shelf losses reward the wrong thing on sparse, bright, or boundary-defined lesions.

    DualEdgeBoundaryLoss, OpticDiscSuppressionLoss, SoftBoundaryLoss, probability-weighted RegionCoherenceLoss, an FOV-masked composite loss, clDice.

What that takes, in practice

Training discipline more than library familiarity. The whole inventory, including what is not known yet, is in the Summa.

Deep learning
PyTorch, timm, segmentation_models_pytorch. Mixed-precision training with GradScaler tuning, EMA, test-time augmentation, k-fold CV with resumable per-fold checkpointing, class-imbalance handling (pos_weight, weighted samplers), threshold calibration on validation only, ordinal regression (CORN), mixture-of-experts, cross-attention fusion, attention gates, ASPP, GeM pooling, deep supervision, frozen-backbone transfer, Grad-CAM.Training discipline more than library familiarity.
Architectures used hands-on
ConvNeXt V2, Swin UNETR V2, DINOv2/v3, EfficientNet, U-Net, U-Net++, Attention U-Net, SwinHRUNetPP, HSMoE-AUNet, image-text encoders.I build, train, modify and evaluate these rather than only reading about them.
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.Most of this is preprocessing nobody writes papers about.
Evaluation
QWK, ICC, AUC-ROC, AUC-PR, Dice, IoU, NMI, lesion-level and image-level recall, sensitivity/specificity tradeoff analysis, per-dataset breakdowns, held-out test discipline and leakage assertions.The per-dataset breakdown is the part that matters.
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.Currently extending this direction toward single-cell foundation models, genomic sequence models and perturbation-based causal analysis.

Research

  1. 2026HierarchiRetinaInterpretable diabetic retinopathy grading from retinal photographsdeposited
  2. 2026Rank Radii Transfer as QuantilesPlacing unseen species inside the taxonomy, not in flat clustersunder review
  3. 2026Automated Renal ReportingThe generator cannot emit a value it did not measurein preparation
  4. 2026HCGT-PGWhat would happen if you knocked the gene downproposal accepted

The Summa

This page is the case. The Summa is the whole of it — a scholastic summa being the form that gathers everything known on a subject into one ordered whole: the prologue, the axioms each project left behind, the full inventory, a textbook written in propositions, and a commonplace book of what has been read.

Open the Summa