sainma
Type “the duel on the lava planet” and the exact clip comes back. Multimodal search that understands what is on screen, what is being said, and what it means — and answers with a watchable clip, not a timestamp.
Organizations sit on hours of video — training libraries, recorded sessions, media archives, product footage — and can retrieve almost none of it. Filename and metadata search cannot answer “find the moment where…”. So people scrub timelines by hand, or the footage goes dark. sainma makes video searchable by what happens on screen.
A natural-language query — a scene description, a line of dialogue, an emotion, an action — is parsed into what to look for across each signal: visual, audio, and text.
Every scene in the corpus is pre-indexed into multimodal vector embeddings that fuse what is visible, what is audible, and what is said. The query runs against all three signals at once, at scene granularity — not per-file.
Five specialized agents — a coordinator, a domain expert, a visual analyst, a dialogue expert, and a clip director — route the query, weigh the candidate scenes, and pick the moment that actually answers it.
Scene boundaries are detected automatically and the answer comes back as a watchable clip — cut cleanly, stabilized, quality-optimized. Results are cached so repeated questions are instant.
Stack: Python agent crew over a FAISS vector index; frame extraction, scene detection, and clip assembly in the media pipeline; smart caching and error-recovery in the coordination layer.
Indexing whole files answers “which video?” sainma detects scene boundaries first and embeds each scene separately, so the index answers the question people actually ask: “which thirty seconds?” Everything downstream inherits that granularity.
Transcript-only search misses what happens silently; vision-only search misses what matters because of what was said. Each scene is embedded across all 3 signals — visual, audio, dialogue — and fused into one vector index, so a query like “the moment she realizes” can land on a scene where nobody says the word “realize.”
One model asked to do everything does everything averagely. sainma splits the judgment five ways — a coordinator that routes, a domain expert, a visual analyst, a dialogue expert, and a clip director that owns the final cut — so each candidate scene is weighed by the agent best equipped to judge its signal.
Our demo instance runs on a movie corpus — type “the duel on the lava planet” and the clip comes back. Film is the hardest test of scene understanding: fast cuts, non-literal dialogue, meaning carried visually. The same pipeline points at any video estate.
Natural-language scene queries over an indexed corpus return watchable clips with cached, reproducible results. Built and operated by us, at our own expense, on our own corpus.
The same pipeline over your footage — training libraries, recorded sessions, brand archives — running in your infrastructure, with the index built once and every question after that answered in seconds.