What

Semantic search inside long-form video. Ask in plain text and get near-exact moments back — a line in a talk, a slide, a take you need. Transcript, frame, and multimodal embeddings fused with weighted Reciprocal Rank Fusion; collections keep each search space focused.

Technicals

Async workers (BullMQ + Valkey) with idempotent jobs, pgvector for embeddings, TUS for long uploads, RustFS object storage, and Vercel's AI Gateway for embedding and transcription.

Why

Drop-in semantic video search as an internal service — precise timestamps without owning the indexing stack, with workers you can scale to your workload.

Where

Upcoming

• Kubernetes and Helm charts for HA deploys
• Multimodal queries — image, video, or speech in, not just text
• Worker stress-testing toward metric-based autoscaling