Ines Dormoy

TEOChat is a vision-language assistant designed for temporal Earth observation imagery. Built with a LLaVA-style architecture (temporally shared vision encoder + LLaMA 2 through an MLP projector), it supports instruction-following over image sequences rather than single frames. The project is trained with TEOChatlas, a 554,071-example temporal instruction dataset spanning dozens of EO tasks, and demonstrates strong performance on temporal reasoning and remote-sensing dialogue benchmarks.

Hugging Face Model · Paper (arXiv)

CongressSignal automates the pipeline from public federal documents to actionable business insights. It crawls GovInfo daily (bills, regulations, executive orders), extracts company mentions and sector-level impact with AI, and delivers personalized email digests based on each user's watchlist. The system combines Python crawling, Supabase/Postgres storage, TypeScript edge functions, and Resend-powered outbound notifications.

GitHub Repository

Winning hackathon project focused on balancing a face-image dataset through GAN-based augmentation. We used a latent-space editing pipeline to generate controlled attribute transformations (for example age, hair, and facial traits), improve class balance, and increase training diversity while preserving identity and visual realism.

The workflow combined custom preprocessing, translation directions, and quality checks to produce submission-ready batches at scale, turning manual augmentation into a reproducible pipeline for the competition.

GitHub Repository