Ashraful Kabir Alif — AI researcher & ML engineer

Open to PhD opportunities & research collaborations

23.81° N · 90.41° E · Dhaka, Bangladesh

I build machine learning that holds up when the signal is faint and the stakes are high.

From JWST coronagraph frames and kidney CT scans to ICU treatment policies and the human genome. Researcher at BRAC University, AI developer at Shahriar Labs — I ship what I study.

Explore the researchSelected builds CV

Studying

MSc CSE · BRAC University

Building

AI @ Shahriar Labs

Local time

Dhaka --:-- GMT+6

Scroll

§01Research

Four threads, one question: can you trust it?

Different domains, same discipline — honest baselines, held-out tests, statistics that survive a second look, and models that say when they are unsure.

01, Astro-ML · BSc thesis

Paper submitted

Finding planets next to blinding stars

Question

Can a model pick out a faint exoplanet in JWST/MIRI coronagraphic images when it has only seen a handful of labelled examples?

Approach: A Siamese ResNet18 trained with contrastive and few-shot learning on direct-imaging frames, separating stars from companions instead of memorising them.
Result: 92.16% accuracy; on exoplanets, segmentation Dice 0.90 and IoU 0.80.

accuracy: 92.16%
exoplanet Dice: 0.90
exoplanet IoU: 0.80

JWST / MIRI
Contrastive learning
Siamese nets
Few-shot

02, Medical imaging

Write-up in progress

Segmenting kidney tumours and stones with a frozen giant

Question

Can a foundation segmentation model be adapted to tiny renal calculi and tumours on CT without full fine-tuning or hand-drawn masks?

Approach: MoE-RenalSAM: a frozen SAM2 Hiera-L encoder with rank-16 LoRA adapters, a class-conditional router over four experts, and weak supervision from contour-guided superpixel pseudo-masks.
Result: Tumour Dice 88.19% (457 slices) and stone Dice 84.06% (276 slices) on a held-out set of 1,749 CT slices, against 68.31% and 29.86% for zero-shot SAM2 — training only 8.0M of 220.7M parameters.

tumour Dice: 88.19%
stone Dice: 84.06%
trainable params: 8.0M

SAM2
LoRA
Mixture-of-experts
Weak supervision

Code

03, Offline RL · Healthcare · CSE753 team project

Study in progress

What a sepsis benchmark can’t see

Question

When an offline RL agent learns sepsis treatment from logged ICU data, what should it do with the actions clinicians rarely took — and can the benchmark tell the difference?

Approach: Action-Masked Conservative Q-Learning on the ICU-Sepsis benchmark: hard logit masking, a Lagrangian safety penalty, and a diffusion behaviour prior guided by a critic ensemble — evaluated exactly, with no Monte Carlo noise.
Result: Diffusion-AM-CQL takes no inadmissible actions and reaches an exact return of 0.794 under both safety rules (simulator optimum: 0.875). Unmasked Deep CQL falls from 0.779 to 0.694 once an inadmissible action ends the episode. Every learned model lands between 0.765 and 0.794, so the benchmark barely separates them; only the simulator’s true dynamics show the gap to 0.875.

inadmissible actions: 0%
Diffusion-AM-CQL exact return: 0.794
simulator optimum: 0.875

Offline RL
CQL
Safe RL
Diffusion policies

Code

04, Genomics · Drug discovery

MSc thesis in progress

Reading variants and molecules with calibrated doubt

Question

Can models of genetic variants and drug molecules attach an uncertainty you can act on, not just a score?

Approach: VariantFusion, my MSc thesis: multi-modal, isoform-aware variant effect prediction with per-gene conformal sets. Alongside it, a GATv2 graph network predicting 23 ADMET endpoints jointly, with MC-dropout uncertainty.
Result: VariantFusion is in progress — no results yet. Separately, the ADMET model reaches test AUC 0.988 on intestinal absorption and 0.892 on blood–brain barrier penetration from a single multi-task backbone.

ADMET endpoints, one model: 23
ADMET · HIA test AUC: 0.988
ADMET · BBB test AUC: 0.892

Variant effect
Conformal prediction
GNNs
ADMET

ADMET code

Side experiments

Smaller studies alongside the main four.

Music genre clustering with audio and lyric VAEs (GTZAN + Bangla)
Audio, lyric and fused VAEs over GTZAN and Bangla song lyrics, scored on six clustering metrics; a β-sweep shows posterior collapse from β ≥ 4.
Code for Music genre clustering with audio and lyric VAEs (GTZAN + Bangla) (opens in a new tab)
Fair & explainable chest X-ray diagnosis
Multi-disease CheXpert classification with Grad-CAM and demographic bias audits.
Monotonic neuro-symbolic semantic loss
Domain constraints as a loss, for tabular models that stay consistent.

§02Builds

Research that ships.

Working software with evaluations you can rerun. Defecto, Tallycam and Citely run entirely in your browser — anything you test them with stays on your device.

01Runs in your browser2026

Defecto

Spots manufacturing defects after seeing only good parts.

Unsupervised visual inspection: a PatchCore-style memory bank over ResNet18 patch features, trained on good parts only, exported to one ONNX graph that runs on WebGPU with a WebAssembly fallback. Thresholds are fitted on validation only; capsules remain the documented weak case (image AUROC 0.708).

pixel AUROC · PCB1: 0.997
image AUROC · PCB1: 0.931
browser-parity images: 306/306

PyTorch
PatchCore
ONNX
WebGPU
React

Live demo Code

02Runs in your browser2026

Tallycam

Detect, track and count objects in video — on your device.

YOLO26n through ONNX Runtime Web plus a two-stage tracker in the spirit of ByteTrack — motion prediction, size-aware gating and class voting; counts directional line crossings and zone occupancy.

track IDs for 8 objects: 14→8
per frame · WebGPU on a laptop iGPU: 210 ms
automated tests: 22

YOLO26n
ONNX Runtime Web
WebGPU
TypeScript
Playwright

Live demo Code

03Runs in your browser2026

Citely

Answers only from your documents — and says no when it can’t.

Grounded document Q&A with section-aware chunking, hybrid BM25 + n-gram retrieval and an IDF-weighted evidence gate that declines off-topic questions. Extractive by design: no LLM, so every answer is quoted from your documents.

right-document citations: 20/20
off-topic declined: 6/6

TypeScript
BM25
Hybrid retrieval
React
Vite

Live demo Code

042026

CaseCraft

Legal drafting that cites its sources and learns from attorney edits.

Three-tier OCR (PDF text → Tesseract → vision LLM), hybrid retrieval at 0.65 semantic + 0.35 BM25 over ChromaDB, citation-mapped drafts, and a diff engine that turns attorney edits into reusable writing patterns.

Drafts carry citations mapped to source passages, plus a grounding report
Per-chunk OCR confidence carried into retrieval

FastAPI
ChromaDB
Gemini
Next.js
Tesseract

Code

More builds

Source & demos

§03Publications

Papers & manuscripts

Statuses are exactly what they say. Venue names appear only once work is accepted.

Showing 5 of 5 papers

2025Published

Pattern Recognition in Disaster Response: Leveraging Machine Learning for Twitter Analysis

Authors: Md. Impreeaj Hossain, Musaib Ibn Habib Mikdad, Sk. Jamil Hossain, Ridwanul Haque, Md. Ashraful Kabir Alif, Ehsanur Rahman Rhythm, Annajiat Alim Rasel

ICITA 2024 · Lecture Notes in Networks and Systems, Springer

Classifying disaster-related tweets with transformer and recurrent models to help responders find signal in social-media noise.

DOI

@inproceedings{hossain2025disaster,
  title     = {Pattern Recognition in Disaster Response: Leveraging Machine Learning for Twitter Analysis},
  author    = {Hossain, Md. Impreeaj and Mikdad, Musaib Ibn Habib and Hossain, Sk. Jamil and Haque, Ridwanul and Alif, Md. Ashraful Kabir and Rhythm, Ehsanur Rahman and Rasel, Annajiat Alim},
  booktitle = {Proceedings of the International Conference on Information Technology and Applications (ICITA 2024)},
  series    = {Lecture Notes in Networks and Systems},
  publisher = {Springer Nature Singapore},
  pages     = {41--53},
  year      = {2025},
  doi       = {10.1007/978-981-96-1758-6_4}
}

2025Submitted
Synergistic Application of Contrastive Learning and Computer Vision Techniques for Direct Imaging-based Exoplanet Detection using JWST Data
Contrastive, few-shot detection of exoplanets in JWST/MIRI coronagraphic images.
2025Submitted
Enhancing Predictive Consistency in Tabular Classification via Monotonic Neuro-Symbolic Semantic Loss
Encoding monotonic domain knowledge as a semantic loss so tabular models stay consistent.
2025Submitted
Explainable Deep Learning for Chromatin Accessibility Prediction
Explainable deep models that predict chromatin accessibility from DNA sequence.
2025In preparation
Fair and Explainable Deep Learning for Multi-Disease Chest X-Ray Diagnosis using CheXpert and Grad-CAM
Multi-label chest X-ray diagnosis audited for demographic bias, explained with Grad-CAM.

Statuses as of Oct 2026 · submitted manuscripts available on request

§04Path

Lab, classroom, and production.

Research and engineering at the same time, on purpose: each keeps the other honest.

2024 — Now
Work
AI Developer
Shahriar Labs — Dhaka
- Built 6+ AI products for education, research and commerce — ReXearch (research networking, 150+ beta users), TryOn (real-time AI cosmetics try-on at 30+ FPS), QuantumSketch (text to 60-second STEM explainers, 50+ made), UniVerse LMS and ProcureFlow.
- Django / FastAPI / PostgreSQL / Redis back ends handling thousands of API calls a day; Dockerised LLM and vision pipelines with 99%+ uptime.
- Django
- FastAPI
- PyTorch
- Docker
Jan 2025 — Now
Work
Software & Systems Analyst
Sirajganj Economic Zone Ltd (SEZL) — Dhaka
- Requirements for an 11-module ERP; Django/SQL prototypes that sped up approval workflows by 30–40%.
- Digitised operations to cut paper-based processes by about 60%; technical lead on the corporate website revamp; Microsoft 365 and database admin for 50+ staff.
- ERP
- Django
- SQL
- BPMN
Ongoing
Research
Researcher
BRAC University · CSE — Advised by Prof. Md. Golam Rabiul Alam
- Exoplanet detection in JWST imaging, fair chest X-ray diagnosis, neuro-symbolic losses and chromatin accessibility prediction.
- Supervises undergraduate peers on research projects.
- Computer vision
- XAI
- Genomics
Spring 2023 — Now
Teaching
Undergraduate Teaching Assistant
BRAC University · CSE
- Python, Data Structures & Algorithms and OOP — supporting 120+ students each semester.
Expected 2027
Education
M.Sc. Computer Science & Engineering
BRAC University
- Machine learning, computer vision and intelligent data systems. Thesis: VariantFusion — variant effect prediction.
2024
CGPA 3.92 / 4.00
Education
B.Sc. Computer Science & Engineering
BRAC University
- Thesis: contrastive learning for detecting exoplanets and stars in astronomical imagery.

Continued in the full CV (opens in a new tab)

Recognition

2025
Orange Corners programme
Top 14 of 265 teams
2021–24
Vice-Chancellor’s & Dean’s Merit Lists
BRAC University
2024
NVIDIA Inception
Startup accepted
2025
Solvio AI Hackathon
Selected participant · National AI Innovation Track
2025
Distinguished Speaker
Advanced Earth & Climate Science Congress · Berlin
2026
Distinguished Speaker
Advanced Chemistry World Congress · Rome

§05About

Every model needs a model card.

Here is mine — what I’m built for, what I’m good at, and a few known quirks.

Portrait of Ashraful Kabir Alif — Ashraful Kabir Alifআশরাফুল কবির আলিফDhaka, Bangladesh

Model card

model: ashraful-kabir-alif

researchengineering

Architecture: Researcher × builder, trained end-to-end
Training data: BRAC University CSE · Dhaka · far too many Colab notebooks
Intended use: Research collaborationsPhD opportunitiesML engineering on hard problems
Strengths: Turning papers into working software
Evaluation you can rerun
Explaining models to non-ML people
Known limitations: Will ask “what’s the held-out split?”
Opens one more browser tab than needed

Now

Oct 2026

Researching: VariantFusion — calibrated variant effect prediction (MSc thesis)
Writing: Safe offline RL for sepsis — what benchmarks can’t see
Shipping: In-browser vision tools on WebGPU

Local time

--:--

GMT+6 · Dhakaঢাকা

Receipts

From Crossref & the CV

peer-reviewed paper (Springer LNNS): 1
manuscripts under review: 3
students supported each term as a TA: 120+
BSc CGPA: 3.92

Stack

23 tools & methods

Learning

PyTorch
TensorFlow
Hugging Face
scikit-learn
PyTorch Geometric

Vision & XAI

SAM2
YOLO
OpenCV
Grad-CAM
SHAP

Decisions & stats

Offline RL
Conformal prediction
MC Dropout
rliable

Shipping

Python
TypeScript
Go
Django
FastAPI
React
PostgreSQL
Docker
ONNX / WebGPU

Motto

Keep learning because knowledge is the key to power!

— GitHub bio

§06Contact

Let’s find the signal together.

Open to PhD opportunities & research collaborations

Open mail app

Elsewhere

AshrafulআশরাফুলKabir Alifকবির আলিফ

Finding planets next to blinding stars

Segmenting kidney tumours and stones with a frozen giant

What a sepsis benchmark can’t see

Reading variants and molecules with calibrated doubt

Side experiments

Music genre clustering with audio and lyric VAEs (GTZAN + Bangla)

Fair & explainable chest X-ray diagnosis

Monotonic neuro-symbolic semantic loss

Defecto

Tallycam

Citely

CaseCraft

More builds

Pattern Recognition in Disaster Response: Leveraging Machine Learning for Twitter Analysis

Synergistic Application of Contrastive Learning and Computer Vision Techniques for Direct Imaging-based Exoplanet Detection using JWST Data

Enhancing Predictive Consistency in Tabular Classification via Monotonic Neuro-Symbolic Semantic Loss

Explainable Deep Learning for Chromatin Accessibility Prediction

Fair and Explainable Deep Learning for Multi-Disease Chest X-Ray Diagnosis using CheXpert and Grad-CAM

AI Developer

Software & Systems Analyst

Researcher

Undergraduate Teaching Assistant

M.Sc. Computer Science & Engineering

B.Sc. Computer Science & Engineering

Recognition

Orange Corners programme

Vice-Chancellor’s & Dean’s Merit Lists

NVIDIA Inception

Solvio AI Hackathon

Distinguished Speaker

Distinguished Speaker

Model card

Now

Local time

Receipts

Stack

Motto

Let’s find the signal together.Let’s find the signal together.

Let’s find the signal together.