OZU is a search engine for what's happening in movies and TV shows. Its AI analyzes entire films: craft, narrative, emotional subtext, plot, etc. This enables searching for something straightforward like "a tense standoff" or "a heist going wrong," to something more nuanced like "someone realizing they've been betrayed" or "a character using humor to deflect pain." You get back the actual scenes, moments that were never searchable until now.
Led the technical direction on the ML side - curating datasets, designing new datasets for new capabilities, identifying promising areas of research to bring into the product, fine-tuning models and embuing them with cinematic intelligence, building custom eval methods for our use cases, connecting to vector DBs, and constantly exploring bleeding edge models and techniques. Wore multiple hats as a co-founder and helped with fundraising, overall product vision, and backend engineering.
Connected in-house models with frontier multimodal LLMs with a novel context engineering approach and long term memory implementation to analyse 2+ hour long videos, detect scene boundaries, and extract various aspects of narrative data like plot significance, character interactions, motivations, set design, editing rhythm, character arcs, etc.
Brought together the best of two worlds: cinematographic expertise captured by CinemaNet, and natural language image ↔ text understanding from CLIP like models. We fine-tuned a hybrid model that produced CLIP embeddings and classifier outputs, learning a robust representation through multi-task training of multiple contrastive and classifier-based objectives across distinct aspects of the frame.