I built an AI-powered video search system that lets users type natural language queries like “climate change interview with the minister” or “sports anchor reacting to a goal” — and instantly discover the most contextually relevant clips. Unlike traditional keyword search, this system understands semantic meaning, enabling editors and journalists to find moments based on ideas rather than exact words.
Tech Stack
ElasticSearch, NextJS, React, TailwindCSS, TypeScript
My Role
The system converts both user queries and video transcripts into dense vector embeddings using a sentence-transformer model. These embeddings are indexed in ElasticSearch, enabling semantic retrieval based on similarity scores rather than keyword matching.
1. Vector search to find the most semantically similar scenes
2. Re-ranking using context relevance and metadata
3. Instant preview generation to help users identify the correct clip quickly
Impact
1. Search latency improved by 80%, enabling near-instant results.
2. Content discovery time dropped by 60%, directly improving editor productivity
3. Became the core discovery layer for multiple downstream AI workflows, including automated tagging and highlight generation