An online platform that teaches young students the fundamentals of AI. Idea to deployed web app in 48 hours with a four-person team (prototyped on Base44): 15+ user interviews, 10+ iterations, 50+ signups, 5 paying users. Since rebuilt from scratch as a full-stack app.
Voice-first real-estate lead engine: a buyer talks to a voice agent for 90 seconds and gets a WhatsApp with matched London listings and a scheduled callback, no human in the loop. I owned the n8n orchestration: trigger on the call, rank listings, send the WhatsApp, log the lead back to Attio.
Downloads and organises all course files from the Canvas LMS in seconds: parallel downloads, PDF merging, a GUI, packaged to run with a single command. Started as a script for a friend; built three times over, each version with a new generation of AI-assisted development.
Building label-efficient 3D segmentation models for automated body composition analysis in CT scans. Extending a prototypical network framework with foundation model encoders (DINO, TAP-CT) for few- and zero-shot segmentation of novel tissue classes without large-scale annotation.
Extended the Structural Causal Bandits framework to operate on equivalence classes (CPDAGs, PAGs) learned through causal discovery. Incorporated non-manipulable variables into the intervention-selection process.
Owned end-to-end development and deployment of an ML-based ETA prediction model for vessel arrivals. Shipped a Telegram-integrated monitoring and alerting system that replaced manual daily health checks. Migrated OpenRPA scrapers to Python (BeautifulSoup, Selenium), lifting success rate from 60% to 99% and processing speed 7x.
Benchmarked supervised and self-supervised deep learning techniques for accelerated MRI reconstruction; contributed to evaluation and reporting of the models.
Built a transfer-time model between properties and major destinations, cutting Google service costs by 30%. Property price estimation models and automated analytics reports.