What I've built
Work
Products shipped, patents granted, research published. The proof behind the writing.
Product · Founder
StockLens
A deterministic financial-intelligence platform that turns fragmented SEC and policy filings into structured, evidence-bearing signal cards. Built for auditability: every number traces back to its source.
Product · In development
Bill-y
An AI-efficiency layer for engineering teams. It splits Claude token spend across the phases of building software (understand, design, build, test, review) so leaders can see whether AI is being used well, not just heavily. FinOps for AI, built as an efficiency coach rather than a surveillance dashboard.
Platform · IoTX · Patent filed
MIS
An AI-native, on-device platform for data-centre cooling. Three layers, adaptive sensing, an agent operating system, and an OS-level decision layer, run the entire closed control loop locally under fixed power and thermal budgets, with no cloud dependency. In pilot with EU data-centre operators under NDA.
A physics-grounded heat-distribution engine steers cooling from a model of how heat forms and moves across the facility rather than reacting to point readings. A deterministic static-DAG execution engine, with a frozen topological order and a backward-BFS activation mask, runs only the active intent's dependency closure, so no inference path is ever unbounded. Safety is architecture: always-active safety and system-health nodes that cannot be masked, a degradation ladder that sheds intent heads before the WCET budget is exceeded, and a per-agent capability envelope that doubles as the security boundary. Extends into HomeX.
Medical device · IoTX · Patent pending
OtoX
An AI-powered medical diagnostic device for ear-infection diagnosis. Rather than a single instrument, OtoX is a complete ecosystem around the device, from capture through to diagnosis, built at IoTX.
Product · Built and shipped
NetworkX Jobs
A jobs and professional-networking platform I designed and built end to end.
Center of Excellence in AI, and before
Projects
Research and engineering builds, from transformers and CNNs to GANs and hackathon systems.
Motive AI Challenge · DETR
Object Detection in Dashcam Images with Transformers
Fine-tuned Meta Research's DETR (DEtection TRansformer, a transformer over a ResNet-50 backbone) on a raw custom dataset to detect cars, trucks, stop-signs and traffic signals. Held up on low-light, occluded and small-object cases by using scene context in each decision. Deployed on a full-stack web app for full-scale testing.
Vision Transformer · Brackish dataset
Marine Object Detection with Transformers
Detected underwater marine life, including several fish species, with a vision transformer. A hard setting of small objects in low light. Fine-tuned on the brackish dataset and reached accurate detection at good confidence.
STRIP AI · Kaggle · SWIN
Blood Clot Origin Classification in Digital Pathology
Classified blood-clot origins in whole-slide digital pathology images with Microsoft Research's SWIN Transformer, a hierarchical transformer built for very high-resolution images. Fine-tuned by transfer learning on a custom dataset, handling gigapixel slides end to end.
Custom CNN · 95% accuracy · Deployed
Vision-WCCS: Water Contamination Classification
Built end to end to learn the full ML lifecycle. Collected the dataset with a 1000x digital microscope in a lab, processed it, and trained a custom CNN, reaching 95% test accuracy. Deployed to the cloud to help communities check contamination levels and reduce waterborne disease.
Network security · Deep learning
Deep-Learning Intrusion Detection for SDN
Built a system to flag malicious traffic reaching a software-defined network controller. SDN's centralized architecture is especially exposed to attacks like DDoS. Developed with Network Security instructor Dr. Talha Naqash to catch attacks in high-speed traffic.
SOFTEC AI Hackathon · Top 10
Patient Readmission Prediction with LSTM
Built an LSTM readmission predictor from scratch over a continuous 24-hour hackathon, finishing top 10 with the highest accuracy on an undisclosed test set, against teams of MS and PhD students.
YOLOv4 · Gradio
Object Detection in Road Images with YOLOv4
Detected cars, buses, trucks and traffic signs with a Darknet pre-trained YOLOv4, served on a Gradio web interface for testing.
NUST Deep Learning Lab
Optical Character Recognition with Tesseract
Built an OCR pipeline on the open-source Tesseract engine to extract text from documents, as part of the NUST Deep Learning Lab document analysis and recognition workshop.
CNN · Personal favourite
Art Generation with Neural Style Transfer
Mapped the style of a Picasso piece onto a photo of my university building using neural style transfer. One of my favourite projects.
1D Conv + GRU
Trigger-Word Detection
Implemented trigger-word detection, the wake-word technology behind Alexa, Google Home, Siri and DuerOS, with an architecture of 1D convolutional layers and gated recurrent units.
DeepLearning.AI
Deep Learning Specialization micro-projects
A set of small builds: EMOJIFY with word-vector representations, improvising a jazz solo with an LSTM, a character-level language model with an RNN, face recognition and art generation with CNNs, pixel-level segmentation with U-Net, and transfer learning with MobileNet v1.
Carnegie Mellon · Open Challenge
Open Catalyst Project
A multi-year effort with an international team to model and discover new catalysts for the climate-energy challenge, building an AI system to replace conventional DFT simulation, which is computationally very expensive.
Collaboration · MIT
Deep Fakes with Generative Adversarial Networks
Worked with an MIT postdoc on a GAN-based deepfake generation system in the early second year of undergrad, to build intuition for how GANs work.
Intellectual property
Patents
Plus 2 PCT applications; OtoX and MIS pending.
Research
Publications
PhD candidate, King's College London. NeurIPS reviewer (x3). Speaker, Stanford AI + Health.
Always learning
Courses & certifications
Off the screen