Projects

A pipeline that generates chest radiograph reports from MIMIC-CXR images, using clinical information from MIMIC-IV.

• The data are split by patient, so no patient appears in both the training and test sets.
• Generated reports are evaluated for factual consistency with the image and with the clinical record.
• Current design: a frozen RAD-DINO-MAIRA-2 image encoder with multi-view fusion and a Q-Former, feeding Qwen2.5-7B fine-tuned with LoRA. Evaluation combines NLP metrics with clinical factuality scores such as CheXbert, RadGraph, GREEN, RaTEScore and RadCliQ.

• Medical AI
• Vision-language models
• MIMIC-CXR
• MIMIC-IV
• LoRA

An urban heat planning tool built by a four-person international team, with a validation partner in Accra, Ghana. I was the technical lead.

• The geospatial pipeline scores 96 grid cells of 500 m, covering 305,510 residents, on heat exposure, population vulnerability and access to healthcare.
• A knapsack budget optimiser selects interventions that reach 11,509 people within a 25,000 USD budget.

• Geospatial
• Optimisation
• Python
• Team lead

An end-to-end EEG pipeline on two public datasets (MUSIN-G and NMED-RP), treating music perception as both classification (emotion and rhythm) and regression (how much a listener enjoyed a song).

• 128-channel EEG was band-pass filtered and cleaned with ICA, then split strictly by participant (15 for training, 5 for testing on MUSIN-G). Models compared included EEGNet, DeepConvNet, CNN-LSTM, a temporal convolutional network and a graph neural network.
• The best model predicted enjoyment with a mean absolute error of about 0.97 on a 1 to 5 scale. CNN-LSTM reached 75.1% accuracy on emotion and rhythm classification.
• Grad-CAM and occlusion analysis showed that predictions depend on sustained patterns over time rather than single events.

• EEG
• Deep learning
• Interpretability
• PyTorch

An accessibility tool that describes an image with BLIP, translates the description with Google Translate and reads it aloud with gTTS, in more than 20 languages.

• End-to-end response time is about 3 seconds, and the interface follows WCAG 2.1 AA guidelines with keyboard and screen reader support.
• The website is live, but its backend is offline because of memory and hosting-cost limits, so image narration only works when you run the code locally from GitHub.

• BLIP
• gTTS
• Accessibility
• Multimodal

A full-stack shipment tracking platform with separate dashboards for admins, operators and drivers. Live GPS locations are pushed over WebSockets instead of polled.

• Role-based access control limits every view and API route to the user's role.
• A React client talks to a Node.js backend through REST and WebSocket services, deployed on Vercel.

• Node.js
• React
• WebSockets
• RBAC

An optical character recognition tool that extracts text from PNG, JPG and JPEG images in English and Hindi using the EasyOCR deep learning library.

• OCR
• EasyOCR
• Python
• Hindi