Welcome to my GitHub!
I'm a final-year Artificial Intelligence & Computer Science student at the University of Birmingham (with a Year in Industry), passionate about building intelligent systems that combine machine learning, deep learning, and performance-oriented design.
- Emotion-aware communication systems (RoBERTa, PyTorch, NLP)
- Predictive maintenance using autoencoders, LSTMs, and causal inference
- Offline, privacy-focused ML tooling (Transformer-based text correction)
- Deployment of real-time deep learning models with Docker, ONNX, and Azure
- Participating in a research project focused on emotion recognition using EEG and eye-tracking data.
- Applying machine learning (PyTorch/TensorFlow) to classify emotional states across cultural contexts.
- Emphasis on temporal modeling, signal preprocessing, and cross-domain generalization.
- Final output includes a research paper/report and presentation.
Real-time emotion detection in Signal messaging
- Developed a dual-model pipeline using RoBERTa for real-time emotion detection in messaging apps.
- Used sequential architecture: binary classification (Emotion/No Emotion) → multi-class (Anger, Fear, Sadness, Disgust, Happiness).
- Incorporated temporal context from the last 5 messages to improve emotion recognition.
- Integrated into Signal messenger with dynamic UI (color-coded message bubbles based on emotion).
- Achieved 91% accuracy (binary) and 88% (multi-class), with >80% user satisfaction in a usability study.
Scalable model to detect and forecast palm oil deforestation
- Used CNNs, Pix2Pix GANs, and K-Means on radar & satellite data
- 98% classification accuracy, forecasted deforestation trends via ARIMA
- Deployed on GCP, published article on Medium
Privacy-first grammar tool using fine-tuned T5-base
- Local execution (no cloud), 5% better than Grammarly on in-house data
- Fine-tuned with ERRANT tagging + WildNLP augmentation
- Deployed in enterprise environment, full write-up on Medium
- Developed a YOLOv8-based model deployed on Android/iOS via ONNX to detect missing safety items in car boots.
- Achieved 98% accuracy and 20ms inference time in real-time conditions.
- Deployed using Azure DevOps in a CI/CD pipeline for continuous model improvement.
Languages: Python, Java, C++, SQL, JS
ML: PyTorch, TensorFlow, Transformers, YOLO, T5
Infra: Docker, Azure, Flask, ONNX, GCP
Libraries: Pandas, OpenCV, ERRANT, HuggingFace