Muhammad
Zeeshan
² ICMA Digital Academy · Instructor, Generative AI
Machine learning engineer working on generative AI, retrieval-augmented systems, and applied deep learning. First-author at IEEE ICASSP 2026. I build RAG pipelines, multi-agent LLM applications, and computer-vision models — and ship them to GCP and AWS.
Peer-reviewed research.
-
Deep Spatio-Temporal Models for Decoding Purkinje Cell Activity in Tongue Movements
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026.Designed deep spatio-temporal neural architectures to decode Purkinje cell firing patterns and map them to tongue motor trajectories — bridging cerebellar neuroscience and signal-processing deep learning.
→ IEEE Xplore -
DR-RAG: Addressing Retrieval Misalignment in Low-Resource Urdu Question Answering
Workshop on Challenges in Processing South Asian Languages (CHiPSAL), LREC-COLING 2026.A retrieval-augmented generation framework that mitigates retrieval drift in low-resource Urdu QA through dense-retrieval alignment and re-ranking strategies.
→ Proceedings
Things I've built.
-
TaxGPT.pk
A multi-agent LLM tax assistant for Pakistan. Specialised GPT-4o-mini agents handle classification, retrieval, computation, and response — each fine-tuned with SFT, chain-of-thought prompting, and instruction tuning. Grounded in ChromaDB-indexed FBR tax law to reduce hallucinations.
-
Academic content augmentation tool
RAG assistant over 400K+ arXiv papers. Improved retrieval precision via multi-perspective query generation and Reciprocal Rank Fusion across hybrid retrievers. Deployed end-to-end on AWS — Docker + ECR, EC2 / Fargate, S3 + CloudFront, Bedrock embeddings, infrastructure as code in Terraform.
-
Chest CT — classification & captioning
A deep-learning pipeline combining ResNet50 (93%+ classification accuracy) with CLIP and GIT to auto-generate radiology-style captions. OpenCV contour analysis localises tumor position and estimates size for clinical context.
-
Urdu news classification
Curated a 2,000+ article Urdu news dataset via web scraping. Fine-tuned and benchmarked BERT, T5, and XLM-RoBERTa for multiclass classification — comparing accuracy and inference cost across model families.
-
SignTalk
Final-year project — an Android app enabling bidirectional sign ↔ natural-language communication using a CNN-based gesture recognition model. Removes the need for a human interpreter for hearing- or speech-impaired users.
Where I've worked.
-
Nov 2025
— PresentInstructor, Generative AI
ICMA Digital Academy
Teach generative AI to CMA students — tooling, LLM fundamentals, prompt engineering, and applied workflows.
-
Jun 2025
— PresentResearch Assistant — S2Cool
LUMS × Northumbria University, UK
Apply ML/DL to HVAC performance optimisation for sustainable cooling. International collaboration on energy efficiency and climate resilience.
-
Jan – Dec
2025Graduate Teaching Assistant
LUMS — CS5316 NLP & CS536 Data Mining (Dr. Asim Karim)
Supported graduate NLP (transformers, LLMs, summarisation, MT) and data-mining courses; mentored students on Hugging Face workflows.
-
Oct – Dec
2024Research Assistant — BAP Lab
LUMS · Bio-Agri-Photonics
Image reconstruction on speckled imaging data using a convolutional autoencoder baseline + HistoGAN with histogram-feature conditioning.
-
Apr 2022
— Jul 2023Microsoft Learn Student Ambassador (Alpha)
Microsoft
Promoted to Alpha tier. Delivered workshops on open source, Git, and AI to 200+ attendees.
-
Jul 2020
— PresentGDSC Lead, then Mentor
Google Developers · FAST-NUCES
1 of 43 GDSC Leads in Pakistan (2020–21). Led 20+ core team and 350+ chapter; organised 15+ events on AI, GCP, and Android with 1,000+ attendees.
Where I studied.
-
Sep 2023
— Jun 2025MS, Computer Science
Lahore University of Management Sciences (LUMS) · Lahore, Pakistan
CGPA 3.41 / 4.0. Coursework and research across machine learning, NLP, data mining, and computer vision.
-
2018
— 2022BS, Computer Science
FAST National University of Computer & Emerging Sciences (NUCES) · Faisalabad, Pakistan
CGPA 3.07 / 4.0. Final-year project: SignTalk — bidirectional sign-language interpretation.
Tools of the trade.
Let's talk.
I'm open to roles in ML engineering and applied research, particularly around LLMs, retrieval, and applied deep learning. The fastest way to reach me is email — I usually reply within a day.