Muhammad Hamza — ML Researcher

01About

I'm Muhammad Hamza — a fourth-year undergraduate at Indian Institute of Technology, Kharagpur, pursuing a Dual Degree (B.Tech + M.Tech) in Chemical Engineering with a specialisation in Artificial Intelligence & Applications. My research interests span

I aspire to work as an ML Engineer and actively seek collaborations in NLP, LLMs, and Computer Vision. Outside the lab, I write on Medium (cricket & sports analytics), have accumulated 1 Million+ views on Quora, and write fiction: I've published a Sherlock Holmes novel and a story collection.

Neural Operators Large Language Models Spiking Neural Networks Multimodal Fusion Computer Vision NLP Theoretical ML Operator Learning

02News

Oct '26

Published two books on Amazon KDP: a Sherlock Holmes novel and a short-story collection.

July '26

Accepted to Algoverse AI as AI Research Fellow, one among only 90 in the cohort from 100,000+ applications.

June '26

Accepted to Amazon ML Summer School 2026

June '26

Preprint on SGD-based optimisation published on arXiv

May '26

Paper on Spiking Neural Networks under review at Neurocomputing (Elsevier)

Apr '26

Paper published at IEEE Xplore

Jan '26

Paper accepted at IEEE ICCECE Conference 2026

Nov '25

Started Research Intern at University of Manchester · Spherical Neural Operators

Oct '25

Paper accepted at iSAI-NLP Conference 2025

May '25

Started Research Intern at University of Oulu, Finland · SNN Pruning & Constrained Optimisation

Apr '25

Started Research Intern at IIM Ranchi · Multimodal ML

Jul '23

Accepted into IIT Kharagpur

03Industry Experience

SWE-ML Intern

Incoming Intern

Glance · InMobi Group

May 2027 – July 2027

Machine Learning Engineering Production ML Systems

Incoming Software Engineering Intern specialising in Machine Learning at Glance, InMobi Group. Will contribute to scalable ML systems and production-grade AI solutions, bridging research-driven models with real-world applications.

04Research

Neural Operators on Spherical Domains for PDE Learning

Research Intern

University of Manchester · Prof. Anirbit Mukherjee

In collaboration with NASA

Nov 2025 – Present

S² manifolds HEALPix ~196K nodes L_max = 511 40× compute reduction

Designed and implemented DeepONet-based neural operators on spherical domains, modeling heat PDE dynamics with L_max = 100+ harmonic modes, achieving stable spectral evolution and accurate eigenvalue decay. Engineered geometry-aware architectures combining Graph CNN (HEALPix, ~196K nodes) & SIREN trunk networks, mitigating spectral bias and enabling high-frequency function reconstruction on S² manifolds. Optimized scalable training pipelines via stochastic spatial sampling (5K/196K nodes per step), reducing computational cost by ~40× while preserving convergence.

Code Blog

Curvature-Weighted Gradient Diversity: A Noise Measure for Geometry-Adaptive SGD

Independent Research

Preprint out on arXiv

Submission to TMLR · Under Review

March 2026 – May 2026

Strongly-convex quadratics Hutchinson estimation 20% lower suboptimality Noise-adaptive schedules

Developed Curvature-Weighted Gradient Diversity (CWGD), a novel noise measure that adjusts for landscape geometry by down-weighting high-curvature parameter directions. Proved that CWGD-modulated cosine annealing reduces asymptotic suboptimality floors by a factor of 1/(1+α), validated by empirical improvements of ~20% over plain cosine annealing across various condition numbers and noise regimes. Identified and resolved estimator degeneracies, delivering a provably exact Hutchinson-based implementation for diagonal Hessian approximations.

Code Preprint Blog

Criticality-Constrained Pruning for Energy-Efficient Spiking Neural Networks

Research Intern

University of Oulu · Prof. Shuai Lee

Mar 2025 – Jun 2025

90% sparsity 95.6% MNIST accuracy 73% energy reduction CQP · CVXPY · snnTorch MNIST · Fashion-MNIST

Designed Criticality-Constrained Quadratic Pruning (CQP), a native PyTorch pipeline fusing surrogate-gradient criticality with weight magnitude to prune SNNs without convex solvers. Formally characterised a continuous-relaxation trap in QP-based pruning causing 44 pp accuracy collapse, and identified a zombie-weight failure mode in Adam fine-tuning. Combined weight sparsification with KL-divergence temporal analysis yields a compound 73% energy reduction at 70% sparsity, retaining 95.6% MNIST accuracy at 90% sparsity. Submitted to Neurocomputing (Elsevier).

Code Blog

Multimodal Fusion for Movie Review Helpfulness Prediction

iSAI-NLP 2025

IIM Ranchi · Prof. Sobhan Sarkar

April 2025 – Aug 2025

Acc 0.875 AUC 0.884 BERT · ViT · CLIP Dempster-Shafer

Developed a scalable multimodal pipeline processing 54,000+ IMDB reviews and 30,000+ frames using BERT, ViT, and CLIP. Pioneered Dempster-Shafer evidential fusion for uncertainty-aware helpfulness prediction, outperforming attention and concatenation baselines across all five industry-standard metrics.

Code Accepted · iSAI-NLP 2025 Blog

05Publications

[1]

iSAI-NLP 2025 IEEE ICCECE 2026 NLP · Vision

Movie Review Helpfulness Prediction using Semantic Alignment and Evidential Fusion-based Multimodal Approach

Muhammad Hamza, et al.

iSAI-NLP 2025 & also accepted at IEEE ICCECE Conference 2026

IEEE Xplore

[2]

Preprint arXiv Optimization in ML

Curvature-Weighted Gradient Diversity: A Noise Measure for Geometry-Adaptive SGD Schedules

Muhammad Hamza & Ayush Goel

On arXiv; submitted to TMLR

arXiv Code

[3]

Under Review Neurocomputing SNNs · Neuromorphic

Criticality-Constrained Iterative Pruning for Energy-Efficient Spiking Neural Networks via Combined Importance Scoring

Muhammad Hamza

Submitted to Neurocomputing — Elsevier (ISSN 0925-2312)

Code

06Projects

ApexSearch: Audio-Visual RAG & Multimodal Video Retrieval Engine

2026

PyTorch & Transformers FAISS Multimodal RAG Streamlit

CLIP & CLAP Late-Fusion Heuristic Groq LPU Inference

Engineered an interactive, end-to-end multimodal retrieval-augmented generation (RAG) system allowing natural language querying and autonomous live commentary over untrimmed, high-speed Formula 1 broadcasts.

Localizes specific events (e.g., crashes, pit stops) in untrimmed sports broadcasts from natural language queries.
Extracts temporal visual features (1 FPS) and synchronized audio chunks, vectorizing them via frozen CLIP and CLAP transformers into independent FAISS databases for sub-50ms retrieval.
Implements a dynamically weighted late-fusion scoring heuristic to resolve cross-modal ambiguity (e.g., trusting audio over visual during heavy motion blur).
Serverless deployment on GitHub and Streamlit Community Cloud, with Vision-Language Model inference offloaded to Groq API calls for zero-cost hosting.

Synthesia: Visuo-Lingual Affective Recommender System

2026

Transformers FAISS Multimodal ML Late-Fusion

ResNet & DistilRoBERTa Dense Vector Search Ablation Study

Engineered a multimodal machine learning pipeline fusing facial expression recognition and textual sentiment analysis to power a dynamic, emotion-driven movie recommendation engine.

Late-Fusion architecture using ResNet (via DeepFace) for visual emotion extraction and DistilRoBERTa for textual sentiment, mathematically blending the distributions.
Dense vector search engine using FAISS and MiniLM embeddings to retrieve and rank recommendations.
Quantitative ablation study showing Late-Fusion outperforms unimodal baselines.
Deployed to a live cloud server, resolving Linux graphics driver dependencies and Python versioning conflicts.

VictorianGPT: Compound AI Conversational Agent

2026

QLoRA RAG ChromaDB Gradio

Qwen 2.5 (3B) 4-bit NF4 Quantization Contextual Prompt Engineering

Engineered an end-to-end pipeline combining Parameter-Efficient Fine-Tuning with Retrieval-Augmented Generation to deploy an atmospheric, 19th-century conversational agent.

Fine-tuned Qwen 2.5 3B Instruct with QLoRA (4-bit) on ~10,000 synthetic conversational pairs to match a Dark Academia and Victorian gothic aesthetic.
Semantic memory pipeline using ChromaDB and MiniLM embeddings to inject historical context from 19th-century literature (Dickens, Stoker, Wilde).
Resolved 'Context Domination' (Lazy RAG) by restructuring the prompt injection hierarchy and tuning repetition penalties to force original prose.
Deployed to Hugging Face Spaces via a custom Gradio interface, managing background threads and memory caching.

ML-Driven Volatility Surface Modeling — NIFTY50 Options

2025

Python Ensemble ML Quantitative Finance

578K+ records 100+ features MSE 0.00009 10× baseline

High-frequency options volatility forecasting pipeline on 578,000+ NIFTY50 index records across 50+ strikes & maturities. Achieved ultra-low MSE of 0.00009 via batch-wise ensemble with strike-aware feature selection.

Engineered 100+ features including Greeks, moneyness, term structure, and microstructure signals for implied volatility surface modelling.
Batch-wise ensemble integrating gradient boosting and regularised regression, outperforming standard ML baselines by over 10× in accuracy.

Competitions & Hackathons

Optiver Trade-athon 2026

Real-time market-making & pricing strategies under dynamic simulated market conditions, among top IIT competitors.

Rank 23Mar 2026 · Optiver

Convolve 4.0 — Pan-IIT AI/ML Hackathon

Screening ML/AI assessment among 3,000+ students from IITs across India.

All India Rank 102Feb 2026 · IIT Madras

Convolve 3.0 — Pan-IIT AI/ML Hackathon

XGBoost loan default prediction · 0.85 precision · 0.91 F1-score · robust class-imbalance handling.

Top 50 / 4000+Jan 2025 · IIT Guwahati