KHAN
AI & Computer Science engineer. I build intelligent systems that perceive emotion, process language, and see the world — from EEG signal pipelines to models deployed in real apps used by real people.
THE PERSON
BEHIND THE MODEL
I'm a final-year AI & Computer Science student at the University of Birmingham, with a year in industry at ALTEN and a research stint at Shanghai Jiao Tong University under Prof. Bao-Liang Lu. I specialise in the deep end of ML (multimodal learning, NLP, affective computing) and I love when research turns into software people actually touch.
Off-screen: sketching fits, dissecting F1 strategy (Ferrari & Merc), re-reading Dune for the third time, dancing, and listening to Kendrick on repeat. He taught me that craft matters – in music, in code, in everything.
WHERE I'VE
WORKED
Built a satellite imagery detection pipeline using CNNs, image-translation GANs, and radar fusion — 95% accuracy detecting palm oil plantations for sustainable land-use monitoring. Developed a privacy-first grammar correction tool beating Grammarly by 5% on domain-specific text. Deployed a real-time YOLOv8 object detection system at 20ms inference with 98% accuracy on Android & iOS.
WHAT I'VE
INVESTIGATED
Ran EEG & eye-tracking experiments on 11 participants for cross-cultural emotion recognition. Designed multimodal ML pipelines: early fusion (SVM), late fusion (Choquet integrals), and deep representation learning (DCCA, Bimodal Deep Autoencoder). DCCA embeddings yielded a 20% F1 improvement over baselines. Authored a 20-page report under Prof. Bao-Liang Lu; workshop submission in progress.
Context-aware dual-stage RoBERTa pipeline for temporal emotion detection in text messages. 91% binary F1 and 88% multi-class F1 on DailyDialog, IEMOCAP & MELD. Integrated into Signal — recolours message bubbles by detected emotion. User study: 80% reported improved emotional clarity. Submitted to AIES 2025.
THINGS I'VE
BUILT
Full-stack AI coaching system that ingests real F1 qualifying telemetry via FastF1 and runs a Temporal Convolutional Network to detect driving mistakes — late braking, oversteer, understeer, missed apex — in near real-time over WebSocket. Uses track distance as the comparison axis (not timestamps) with DTW alignment, enabling principled cross-driver comparison. Trained on Hamilton & Verstappen data across 10 circuits. ONNX export hits ~8ms CPU inference. React dashboard streams live telemetry, a track map, corner-by-corner breakdown, and an optional Ollama post-lap debrief.
PPO agent trained to autonomously dock and station-keep a spacecraft using real Clohessy-Wiltshire equations in a custom Gymnasium environment. Full 2D and 3D implementations with RK4 integration. 100% docking success in 2D, 99% in 3D. Station-keeping scores >99.7% of theoretical maximum. Non-obvious reward engineering required to solve the hovering problem, initialisation problem, and CW secular drift.
A real-time, dual-model emotion detection system integrated into the open-source Signal messaging app. A binary RoBERTa classifier detects emotional presence; a multi-class model identifies the specific emotion — both using temporal context from the last 5 messages. Chat bubbles recolour dynamically to reflect emotional tone. Trained on DailyDialog, IEMOCAP & MELD. Evaluated with 15–20 participants: 80%+ reported improved emotional clarity in conversations.
Dual-head Transformer that forecasts 21-day realised volatility and classifies the current market regime (Low / Normal / High Stress), then injects that forecast into a Monte Carlo simulation engine to generate realistic price paths. Replaces the fixed σ in standard Geometric Brownian Motion with a time-varying σ(t) — walk-forward validated on 11 tickers including SPY, AAPL, BTC-USD. Live Streamlit dashboard with VaR, CVaR, and percentile fan charts.
Multimodal pipeline fusing EEG & eye-tracking data for cross-cultural emotion recognition. DCCA embeddings outperformed all baselines by 20% F1. 11-participant study at SJTU under Prof. Bao-Liang Lu. Workshop submission in progress.
Scans a physical cube via webcam using HSV color detection on a fixed 3×3 grid overlay — robust to lighting variation where RGB fails. Solves it with the Kociemba two-phase algorithm (typically under 20 moves, avg 8.1ms). Visualises the full solution as a step-through interactive 3D cube rendered in OpenGL.
Webots simulation of a socially-aware care robot for dementia patients. Custom-built SLAM pipeline (LiDAR + wheel encoders + IMU complementary filter), frontier-based autonomous exploration, and a novel Social A* planner with Gaussian proxemic cost fields. Three social profiles keep the robot at a safe, comfortable distance from the patient.
CNN + image-translation GAN + radar fusion pipeline that identifies palm oil plantations from satellite imagery with 95% accuracy. Built at ALTEN to support sustainable land-use monitoring.
Solving airline crew scheduling as a Set Partitioning Problem using three metaheuristic algorithms. The Improved Binary Genetic Algorithm with stochastic ranking converged to the known optimum on every run for two of three benchmark datasets, and in 27/30 runs on the third — significantly outperforming standard SA and BGA.
MY
TOOLKIT
THINGS I'M
PROUD OF
Google DeepMind Scholarship — MSc Artificial Intelligence
Competitive full scholarship to study an AI Masters. Starting Sep 2026, graduating 2027.
Conference Submission — AIES 2025 (AAAI/ACM)
Temporal Emotion Detection in Text Messaging via Dual RoBERTa Models. Under revision.
First-Class Final Year Project — EmoSMS
Graded at First-Class level by the University of Birmingham.
Invited Research Position — SJTU Shanghai
Selected for summer research at the Center for Brain-Like Computing & Machine Intelligence under Prof. Bao-Liang Lu.
Technical Writer — Medium
Two published articles on Transformers for grammar correction and AI for sustainability.
WHERE I
STUDIED
Awarded a competitive Google DeepMind Scholarship to pursue a Masters in AI. Starting Sep 2026, graduating 2027.
Final Year Project graded First-Class: "Temporal Emotion Detection in Text Messaging via Dual RoBERTa Models." Submitted to AIES 2025. Year in Industry at ALTEN LTD (2022–2023). Research exchange at SJTU Shanghai (2025).
THE
ROAD SO FAR
Scroll horizontally →