Skip to main content
Computer Vision · Video-to-Data

Vision

Turns camera feeds into structured, queryable data — video and images in, normalised events out, ready for alerting today and ML tomorrow.

Overview

Vision converts raw video and image streams into structured, normalised data. It runs platform-side AI inference and ingests camera-native detections from smart devices, converging both into one event model with every record linked back to its source frame as evidence. That structured data is the deliverable — notifications are one way to consume it today, and the same normalised events are the foundation for feature generation and ML models downstream.

Outcome

Running as Sapium's internal pilot: camera feeds across multiple use cases, including real-time people counting, are continuously converted into normalised, evidence-linked event data, with email alerting live as the first downstream consumer. Pilot success is measured over a 60-day window by growth in active use cases and connected cameras ahead of wider rollout and additional downstream applications.

What it does

Vision ingests two kinds of input side by side: raw frames from cameras that get run through platform-side AI inference, and structured detection events parsed directly from smart cameras (Dahua-class devices) that already do their own on-device processing. Both paths converge into the same normalised event model — the output is structured data, not video.

Every event carries a URI reference to its source frame instead of embedding the image, so evidence stays queryable and replayable without bloating the transport layer. A dedicated persistence module keeps a durable, retention-managed history (30 days by default, configurable per deployment), turning what used to be a video archive into a queryable dataset.

Where it fits

Vision is the visual counterpart to Conduit: where Conduit turns raw telco data into structured features, Vision turns raw video and image data into structured events. Notifications are the first thing built on top of that data — when a normalised event matches a configured condition, an alert goes out — but the notification pipeline is one consumer among others the architecture is built for. The same evidence-linked, normalised event stream is what feature generation and ML models downstream would draw on.

A queue-managed backbone sits between camera ingestion, AI/parsing, event persistence, and notification delivery, decoupling each stage so new consumers of that data can be added without redesigning ingestion or inference. core_mgmt is the orchestration and API boundary; it exposes the structured event data without callers needing to reach into the queue or the AI engine directly.

Built to extend beyond alerting

Camera-native and platform-AI use cases (including real-time people counting) run side by side on the same normalised model, so adding a new detection type doesn’t mean redesigning the pipeline. Every record traces back to a source frame with a linked evidence URI, not a bare event — which is what makes the data trustworthy enough to build a second or third consumer on top of, not just an alert.

Ready to move from data to decisions?

We work with mid-tier telecom operators ready to take AI from proof-of-concept to production.

Start a conversation