About Overshoot and real-time visual intelligence
Overshoot builds a real-time vision API for developers. The product turns live WebRTC video into a queryable Stream and lets applications ask vision-language models about current or recent visual context through an OpenAI-compatible interface. Overshoot runs cloud-hosted inference and retains 600 seconds of Stream history for frame and segment references. The company focuses on the infrastructure between live media and an application decision. Customers build the surrounding product, including source management, prompts, validation, workflows, interfaces, durable storage, access, and human review.
- Product
- Vision APILive Streams plus cloud-hosted vision-language model inference.
- Response
- 200msTypical response time for Overshoot-hosted vision models.
- Live transport
- WebRTCLiveKit publishing from browser, native, or server code.
- API
- /v1betaStreams, models, and OpenAI-compatible chat completions.
- Regions
- 2us-west1 and us-central1 are documented.
What Overshoot builds
Overshoot packages live media transport, Stream lifecycle, recent visual memory, temporal selection, model routing, and streamed output behind a developer API. It serves applications that ask changing language questions about video while events remain recent, including video agents, operator assistance, visual review, and interactive camera features. The service handles cloud-hosted VLM inference. Your application decides when to query, how to validate the answer, and what workflow follows.
How the product works
POST /v1beta/streams returns a LiveKit room and a 300-second lease. Browser, native, or server code publishes WebRTC video, and the active Stream retains 600 seconds. A chat completion references the latest frame, an exact moment, or a segment through ovs://. Requests may also use HTTPS or data URLs. /models reports availability, SSE streams output, hosted models respond in 200ms, and documented regions are us-west1 and us-central1.
- Create a Stream
- Publish WebRTC video
- Select current or recent context
- Run a hosted VLM
- Stream output to the application
Where the product boundary sits
Developers own scheduling, prompts, media windows, validation, tools, review, source management, identity, authorization, long-term recording, evidence retention, interfaces, and final actions. Overshoot does not run an autonomous loop on every frame. The public API has no direct RTSP, RTMP, ONVIF, USB, PLC, or SCADA integration and provides no edge runtime or named hardware support. Offline, local-only, hard real-time, and safety-rated functions require appropriate systems outside this cloud VLM path. Connected products can combine an existing detector or event source with a focused cloud VLM query.
How to evaluate Overshoot
Start with one representative source and one bounded question. Build a labeled set with unknown and failure cases, run the full Stream lifecycle, and measure task correctness, first token, completion, errors, and reviewer effort under realistic load. Record model, region, frame count, prompt, output limit, and percentile. Include publisher recovery and model unavailability in the same proof. The published Qwen and Gemma benchmarks are workload-specific engineering evidence. Confirm current limits and commercial requirements directly; this page makes no customer-outcome, uptime, SLA, free-tier, or capacity claim.
Frequently asked questions
Is Overshoot an edge runtime?
No. Overshoot runs VLM inference in managed cloud regions. Devices publish video over WebRTC.
Does Overshoot provide a complete camera system?
No. It provides live Stream and VLM API infrastructure for software teams building their own product and workflows.
References
See the product on live video
Try one visual question in the playground, then build the complete Stream lifecycle with the quickstart.