Open to research & collaboration

SoyabulIslam Lincoln.

AI engineer for medical image processing. I build segmentation and computer vision models in Python, and write custom GPU kernels when they need to run fast.

See the work Walk the Great Wall Email
Teton Private Ltd.ML Engineer · R&D
11 papers9 published
Dhaka, BDGMT+6
Soyabul Islam Lincoln
01

Spotlight

Two builds
01 · Flight & capturePipeline
Illustration · images → orthomosaic → DSM → dense cloud
01Computer vision · GPU kernels · Python

pyOrthomosaic

Drone photos in, survey-grade maps out, natively in Python.

Orthomosaics, DSM / DTM elevation models, dense point clouds and 3D meshes from a folder of drone images. No Docker, no VM, no OpenCV, GDAL or PROJ: the heavy lifting runs in Cython-compiled C that uses every CPU core, and picks up NVIDIA (CUDA) or Apple-silicon (Metal) GPUs when they are there.

  • Two pipelines: a fast planar orthomosaic, and a 2.5D / 3D build with a true orthophoto on the DSM
  • DSM, DTM, nDSM, classified LAZ / PLY clouds, OBJ / GLB meshes, contours and a quality report
  • Bundle adjustment with RTK priors and ground control points, rolling-shutter correction, AI sky masks
  • Five native C kernels; pip install pyOrthomosaic runs on CPU with only NumPy and Pillow
PythonCython / CCUDAMetalGIS
View on GitHub Details
Normal sinus · 72 bpmSynthetic · Lead II
Illustration · 25 mm/s · Gaussian beat model, as in the project
02Biomedical signals · DSP

ECG SmartSim

An ECG synthesis and signal-processing workbench in the browser.

Flask and NumPy / SciPy on the server, a vanilla-JavaScript Canvas display in front. It synthesises 23 cardiac conditions, injects realistic noise, and lets you chain filters and analyses that update live. Atrial and ventricular events run on separate timelines, so blocked P waves, Wenckebach progression and AV dissociation come out of the model rather than being drawn by hand.

  • Sinus rhythms, AF / flutter, PVCs, VT, AV blocks, bundle-branch blocks, WPW, STEMI territories, hyperkalaemia
  • Lead II or an illustrative 12-lead projection with editable P / Q / R / S / T components
  • Up to eight chained stages: Butterworth, Chebyshev, elliptic, FIR, notch, Savitzky-Golay, baseline removal
  • FFT, direct DFT, Welch PSD, STFT, autocorrelation; QRS detection with RR, SDNN and RMSSD
PythonFlaskSciPyJavaScriptDocker
View on GitHub Details
02

From signals to stories

Bio · five chapters

I came into machine learning through electronics, where you learn to distrust a signal that looks too clean. Now I build models for medical images and body signals: chest X-rays, CT and MRI scans, ECG, EEG. Most weeks that means cleaning messy data, restarting training runs, and reading papers until something finally clicks.

What I care about is a model a clinician could actually trust. Not just the prediction, but the reason behind it. That's the thread running through the five chapters.

Vision: making the model show its work

Medical images arrive noisy and thinly labelled. I've worked on detection in chest X-rays, segmentation of anatomical regions, and sharpening blood-cell microscopy with a super-resolution GAN before classifying it.

The part I keep coming back to is explanation. A heat map that lands on the lung field is worth more to me than a tenth of a point of accuracy, so almost everything I build gets checked with Grad-CAM or a similar method before I trust it. A model that has quietly learned to read the corner of the scan instead of the anatomy is a bug, and I'd rather find it on a Tuesday than in a clinic.

Signals: the body's own wiring

ECG and EEG are where I've spent the most time. With EXGnet the aim was a single-lead classifier a small clinic could actually run: accurate enough to be useful, light enough for a Raspberry Pi. To get there the model had to learn where to look, so training was supervised with a cross-correlation loss tied to the clinically relevant parts of the beat instead of letting attention wander.

CARDxnosis goes further and writes the report: anatomical lead grouping, cross-attention between the raw waveform and hand-crafted descriptors, and a language model turning the output into something a cardiologist can read. The hard parts were rarely architectural. They were things like keeping a model honest when two annotators disagree about the same strip of rhythm.

Brains, drawn as graphs

Some problems don't fit an image. For the connectome work I treat a brain as a graph: regions as nodes, connectivity as edges, then let a graph network classify it. On ABIDE and ADHD-200 the fusion model finished three to five points ahead of the baselines we compared it with, which on this kind of data is a real gap.

It's also the work I find hardest to explain to people outside it. Saying the edges between two regions are behaving differently takes longer than pointing at a highlighted scan. I'm still working on that.

Text: when the data is language

Text started as a side road for me, a toxic-comment classifier, and became something I use constantly. Clinical notes are dense, inconsistent, and often the only place the useful detail lives, so I've built summarisers and report generators that turn a model's output into sentences someone can act on.

It's also the discipline that keeps me humble. Language models are very good at sounding certain, and in a medical setting that's the last thing anyone needs.

Shipping: getting it out of the notebook

Research code and something a hospital can run are not the same object. A lot of my time goes into that gap: quantising models so they fit on an edge device, wiring Flask and FastAPI services that stay up, and sitting in calls with clinicians in Dhaka, Delhi and Tokyo while they tell me the interface is wrong.

That loop is my favourite part of the job. The model barely changes; everything around it changes a lot.

Why I keep doing this

Biomedical AI is the corner I keep choosing. The data is guarded, the timelines are long, and a mistake can cost someone something real, which is exactly why it holds my attention. Most of the job is translation: between a clinician who describes a problem by symptoms and an engineer who describes it in tensors. I'm useful in both rooms.

03

Now

July 2023 to present

Machine Learning Engineer (Tier I), R&D

Dhaka, BD

Teton Private Ltd. · July 2023 to Present

Driving innovation in R&D by developing advanced machine learning models for biomedical applications and beyond.

Full timeline

5+Years experience
20+Projects completed
10+Research papers
7+Internships
04

Selected research

Eleven papers · nine published

All publications arXiv preprints

05

More projects

Six of seventeen

MSA-TransUNet++

01

Hybrid CNN and transformer segmenting the stomach, small bowel and large bowel in abdominal MRI.

SegmentationSwin TransformerMRI

Inception Swin-UNet

02

Micro-CT segmentation of void, grain and calcite, with focal loss and a TFLite export.

SegmentationSwin-UNetTFLite

MGTKAN

03

Kolmogorov-Arnold networks for tabular data with graph feature fusion and knowledge distillation.

KANGraph networksDistillation

Contextual Predictive Maintenance

04

Job context beats KPI telemetry for predicting cluster failures: AUPRC about 0.70 against 0.46.

LightGBMVAEGCP

LLM Competency Extraction

05

Llama 3.2 over 4.9 million job postings, mapped to O*NET and ESCO in a skills knowledge graph.

LLMKnowledge graphNLP

Pneumonia Detection

06

Analyzed chest radiogram images to detect pneumonia using image processing and machine learning.

Medical imagingOpenCVDeep learning

All projects

06 · The 3D portfolio

Walk the Great Wall as a panda.

The same portfolio, built as a mountain ridge you can explore in the browser. Every watchtower holds part of it; light the beacon at the end to get in touch.

  • 10 watchtowers
  • 17 projects
  • 11 papers on stone steles
  • 3 golden bamboo hidden
Start walking
07 · Contact

If your work needs strong engineering, deep learning expertise and a feel for biomedical complexity, I am always open to collaborate.

islamsoyabul98@gmail.com Contact page GitHub