Enterprise live video analytics with retained context

Place the API inside an owned architecture

A typical application has a publisher layer, Stream registry, inference scheduler, prompt catalog, response validator, workflow service, and review interface. Overshoot sits between the publisher and inference client. Existing identity systems determine who can access each source. Existing storage and event systems hold durable records. The application maps a business event to a specific Stream and visual anchor.

Keep these boundaries visible during evaluation. The service does not include a camera fleet, video-management dashboard, identity provider, event bus, or long-term recorder. Avoid assuming direct support for RTSP, RTMP, or ONVIF. A gateway can publish a LiveKit track after your team validates and operates that conversion path.

Design regional Stream ownership

Overshoot documents us-west1 and us-central1. Assign each Stream to the selected region and keep Stream references in a single multi-source request within one region. Store region next to the Stream id so request routing is deterministic. Evaluate publisher location, user location, internal policy, and measured network time when choosing a region.

The Stream lease is 300 seconds. Centralize keepalive ownership so duplicate workers do not create confusing lifecycle behavior. Each successful keepalive returns a fresh publish token. On expiration, create a new Stream and update the registry. Delete during planned teardown. Monitor active state, frame recency, retained counts, and ended reasons as separate operational signals.

Standardize prompts and response contracts

Treat prompts as versioned product assets. Give each capability a purpose, allowed sources, media-selection rule, model policy, output schema, unknown behavior, and acceptance set. Keep prompts narrow enough for independent evaluation. Use response_format only with models whose support has been tested. Validate all fields before publishing an event or invoking a workflow.

Discover ready models through /v1beta/models. Define whether a capability permits a tested alternate model because models can differ in output and latency. Record model id, prompt version, media anchor, and response with the workflow event. If sensitive output should not enter general logs, emit structured operational metadata and keep review data in an access-controlled store.

Build observable performance tests

Overshoot-hosted models respond in 200ms. Enterprise acceptance should measure the complete path at realistic source count and request bursts. Record time to first token, full response, inter-token latency, errors, achieved throughput, model, region, resolution, frame count, max_fps, prompt size, and output budget. Report percentile and test duration.

The Qwen and Gemma engineering results are workload-specific. The Qwen processor reached 81 ms p90 for 15 480p frames on one H200 after optimization and stayed near 80 ms at 5 QPS. The Gemma test reached about 120 ms p95 TTFT at 20 QPS for six 480p frames after batching. Use these as methodology examples, not enterprise capacity claims.

Set governance before broad source access

Map sources to allowed purposes and users. Apply least privilege to API keys, Stream mappings, prompts, answers, and stored evidence. Define retention independently from the 600-second Stream history. Decide which visual content can be sent, which fields can be logged, and when a person must review an answer. Document model limitations in the user experience.

Keep consequential and safety-critical decisions behind deterministic policy and appropriate review. Add an unknown state and a visible unavailable state. Test prompt injection present inside visual scenes if downstream actions depend on model text. Separate observations from tool instructions. An enterprise system needs a policy boundary that a visual message cannot rewrite.

Run a technical and operational evaluation

Choose one workflow with representative media, expected request rate, region, and review path. Validate API behavior, model quality, network path, keepalive, reconnects, expired Streams, unavailable models, partial SSE, and parsing. Confirm current limits and commercial terms directly. Product copy and benchmarks do not establish uptime, SLA, quotas, or pricing.

Run shadow mode with the people who own the current process. Measure task correctness and reviewer effort. Review security and data handling. Expand only after the workflow meets a written threshold and has an operational owner. This process produces evidence specific to the enterprise environment without relying on invented customer outcomes or unsupported service guarantees.

References

• Regions
• API limits
• Contact Overshoot