I'm Tasawar Abbas Khan — I build LLM, RAG and computer-vision systems for medicine and other high-stakes work, then make them show why they decided what they decided. Published in IEEE Xplore. Every system below is live — you can try it right now.
My name is Tasawar Abbas Khan, an AI researcher and machine learning engineer from Islamabad. I teach it by day, publish it, and ship it — by the way, most people just call me Tasawar!
I work where AI meets consequences: medicine, diagnostics, decisions people can't afford to get wrong. Getting a model to produce an answer is the easy part. Getting it to show its reasoning — so a doctor can agree or overrule it — is the part I care about.
That question is what became FIRE, my framework published in IEEE Xplore: it turns a deep network's prediction into rules a human can actually read. The same instinct runs through everything else I build — grounded in real evidence, never a confident guess.
I don't hand over black boxes, and I don't ask you to take my word for it. Every system I've built is deployed in public, so you can open it and test it yourself.
Rather see than read? Try a live demo — or just say hello.
Or, keep scrolling if you want to learn more…
Built for domains where a wrong answer has a cost.
I take an idea from literature review through architecture, training and evaluation — and then actually deploy it, so it survives contact with real users.
Multi-agent LLM applications grounded in real sources — vector search with FAISS, LangChain orchestration, and retrieval that stops models inventing answers.
Deep models that justify themselves — Grad-CAM heatmaps, SHAP attributions and rule extraction that turn a prediction into evidence a clinician can weigh.
CNN, ResNet, EfficientNet, U-Net and YOLO pipelines for classification, segmentation and detection — trained, evaluated and shipped behind a real API.
Every project below is deployed and running on Hugging Face. Open one, upload something, and watch it work.
A multi-agent medical AI that queries the live PubMed API across 5M+ papers to build evidence-grounded differential diagnoses. FAISS vector embeddings keep every claim tied to a real source instead of a hallucination, and integrated vision-language models read X-rays and MRI scans.
A bilingual (English/Urdu) multi-agent RAG application for analysing skin conditions, backed by a custom EfficientNet-B0 trained in PyTorch. Grad-CAM heatmaps show exactly which region of the lesion drove the prediction, and a Pakistan-specific medicine database grounds the recommendations locally.
A full-stack semantic matching engine that scores a résumé against a job description in under 30 seconds. A FastAPI backend runs cosine-similarity scoring over Sentence Transformer embeddings to surface skill gaps, then generates an ATS-ready rewrite, a cover letter and interview prep.
Upload a PDF or DOCX — or paste raw text — and get a sentence-level breakdown of how much reads as AI-generated, a plagiarism similarity score with matches, and a "humanize" pass that rephrases flagged passages in a simple or professional register.
Deep learning models predict diabetes well, but their black-box nature costs them the trust of the clinicians who have to act on the output. FIRE pairs the accuracy of a Deep Neural Network with the interpretability of an XGBoost surrogate: polynomial feature transformations and attention mechanisms capture nonlinear interactions among features, while SHAP values translate the DNN's reasoning into human-readable "if-then" rules. Crucially, those rules are converted back out of normalised feature space into real clinical units — mg/dL, mmHg, kg/m² — so a doctor reads them in the language they already work in.
Currently a PhD researcher in Computer Science (AI) at COMSATS University Islamabad — ongoing work on explainability and trustworthy AI.
Articles are published from the admin panel — the newest three appear here automatically.
Ideal if you need an ML engineer who can read the literature, build the model, and take it all the way to a deployed, explainable system.
Ideal if you're working on explainability, medical AI or trustworthy ML and want a co-author who also ships working code.
Ideal if you have a defined AI problem — a RAG assistant, a vision pipeline, a model that needs to be explainable — and want it built properly.
Something that doesn't fit neatly into a box? Just email me →
Still wondering whether AI can be trusted with the decisions that matter?
Or just say hello — I reply to everything.
