physical-world media · self-hosted

vidarax_

Vidarax turns continuous video into actionable, replayable assertions. Give it a live stream or recording. For recorded audio-video, it can keep local transcript moments on the source timeline so your application can move from a spoken phrase to the matching video interval without owning a media pipeline.

public alpha self-hosted documentation audio + transcripts source on GitHub read the system tour write for early access

every frame is checked. selected moments become durable, queryable events.

How it works

ingest

Upload a recording or attach WebRTC, RTSP, or HLS. Recorded audio and video stay on the source timeline. Every accepted live session first reserves a bounded memory and worker envelope, then enters one supervised processing generation.

decide what matters

A deterministic per-frame filter checks scene cuts, flicker, ghosting, exposure shifts, and loops before inference. Optional novelty signals can reduce repeated analysis. The engine reports every selection decision and its timing so each deployment can measure the savings on its own streams.

turn it into action

Selected frames reach the configured vision provider. When recorded analysis enables local audio, its MP4 and WAV windows cover the same source interval and speech recognition adds timestamped transcript moments. A configured audio-video provider can relate those moments to visible activity. Trigger programs and delivery hooks consume the result.

Selectivity and end-to-end throughput depend on the stream, provider, and hardware. Vidarax exposes the decisions and timings needed to measure that deployment directly. Recorded transcript moments use source-relative audio timestamps, so applications can seek back to the matching video interval without reconstructing the timeline. One failed stateful stage faults the live generation as a unit. It is never restarted underneath the decoder and temporal state that came before it.

Built for operators

Own the media plane. The server, event log, and selected keyframes run on infrastructure you operate.
Choose where inference runs. Keep selected frames local with an OpenAI-compatible provider, or explicitly configure Gemini.
Bound peak live work. Admission reserves queue, pool, scratch, sidecar, and worker capacity before a live run becomes durable.
Fail as one generation. A supervisor owns every stateful stage, closes the peer on an unexpected exit, and stops and joins the sibling set.
Act on ordered records. Cursor-based SSE replays after reconnect, signed webhooks retry delivery, and stable event IDs make duplicate handling explicit.
Keep media binary. Images live in a content-addressed sidecar. Events carry references, hashes, types, and byte counts, never JSON/base64 payloads.
Keep speech on the media timeline. When local audio is enabled, recorded analysis emits timestamped transcript moments. A configured audio-video provider can relate them to visible activity, while the retained media stays binary.
Run bounded rules. Versioned trigger programs compile, validate, and replay through the REST API, CLI, and TypeScript SDK before they are attached to a live generation.
Stage edge releases. Signed model manifests advance through shadow and canary checks before activation, with monotonic versions and rollback on failure.
Keep spatial context. Frame events record source dimensions, normalized crop, exact pixel region, and analyzed extent under the vidarax.image.v1 coordinate contract.
See the pressure and faults. The dashboard derives healthy, idle, saturated, or faulted state from backend metrics, including capacity rejections, worker exits, and clean versus forced shutdowns.

where it is aimed: fleet cameras facility monitoring loading-bay operations robotics observation recorded session review

Early access

Vidarax is a public alpha. The documentation covers the API, ingest paths, synchronized audio and transcripts, trigger programs, live delivery, edge deployment, and the internal media plane. For the full argument, read A video stream through Vidarax. If you want to test it against your streams, write to cosmin@bararu.com and say what the streams are and what events you need out of them.