Projects

Get immediate feedback about your personal carbon footprint when sharing pictures of food on social media. A Python backend monitors the user's Instagram profile for new posts, fetches the image, and analyses it with a deep learning model that classifies the meal, detects ingredients, and generates a recipe. Ingredients are matched to their estimated carbon footprints; high-carbon ingredients trigger suggestions for lower-impact alternatives and an optimised recipe. The user is notified via Firebase Cloud Messaging.

An intuitive app that helps users counteract an upcoming cold infection using an early-detection algorithm based on enriched location data and resting heart rate (RHR). Scientific studies show that RHR during the first 4 hours of sleep can indicate whether the immune system is already fighting influenza — before the person notices. Orange combines this with statistical influenza density data from the Robert Koch Institute and the user's historic location data to determine exposure to high-risk environments, then provides personalised counter-measure recommendations (zinc, rest, stay home).

Predictive maintenance for your body. Line connects to a Garmin smart watch, collects daily activity data, and builds an individual injury risk profile. It calculates whether current training is too intense or miscontextualised, then suggests personalised preventive exercises. To ensure exercises are performed correctly, Line uses a segmentation model to crop the body from video and a pose estimation model to extract skeletal keypoints, providing real-time movement feedback (e.g. squat form detection).

Personalised insurance through social media footprint. Unique analyses a user's social media presence to generate a personalised insurance profile, offering tailored coverage and pricing. Built as a full-stack web application with a Django backend and a React/Relay/GraphQL frontend.

Frictionless payment app with a deep-learning-based intelligent text-to-image search (built in 2016, ahead of its time). The Android frontend provides a seamless payment UX, while the Python backend powers the intelligent search and payment logic. Won the grand final of the POST /bank Hackathon Roadshow 2016/2017.

Drop in a photo, get a numbered, ready-to-paint template with a palette and per-region colours — in your browser, in about a second. The pipeline runs entirely on your own machine or server: no third-party upload, no SaaS lock-in, no tracking. A FastAPI backend runs the image-processing pipeline (k-means clustering in perceptually-uniform colour spaces, connected-component region extraction, pixel-exact outline tracing, largest-inscribed-circle number placement) and a vanilla-JS SPA with an interactive painting canvas serves as the frontend — no build step, no framework. Hardened with CSP, non-root containers, readOnlyRootFilesystem, sliding-window rate limiting, and security headers. Production-deployed via Docker Compose and Kubernetes with Traefik + cert-manager (Let's Encrypt).

Listen to any scientific paper — sentence by sentence, with live word-by-word highlighting. Paste an arXiv link, any article URL, or open a local PDF. Parses the document, converts inline LaTeX math to spoken English, synthesises natural neural speech (Kokoro-82M TTS), and highlights the spoken word in real time. Production-deployed with Kubernetes, nginx, and TLS via Let's Encrypt.

A self-directed, closed-loop research system that ideates, implements, trains, evaluates, and iterates on ML experiments without human intervention. Features 8 specialised expert personas (hyperparameter, architecture, loss, data, GPU, optimizer, high-risk, web-research), HPC cluster integration (SLURM/LSF), git-native experiment isolation, W&B tracking, and live human-in-the-loop steering via user_prompt.md.

An AWS-hosted web app accompanying a TV-show tippspiel (betting prediction game) used by 30+ people in the department. Full-stack application with a Python backend, TypeScript frontend, and live score tracking.

A framework for data augmentation for 2D and 3D image classification and segmentation. Widely used in the medical-imaging PyTorch ecosystem.

An intuitive PyTorch-centric package to streamline setup and visualization for ML experiments. Integrated with TensorBoard and Visdom.