Keep recording off the critical network path.
Capture, indexing, retention, and core detection run on the edge. The cloud can extend access and operations without becoming a prerequisite for the archive.
Trust & architecture
Resilience is a system property, not a slogan.
SurveillX keeps critical capture close to the source, separates media and control concerns, exposes operational health, and describes capability boundaries plainly.
Engineering principles
Designed for failure, inspectability, and control.
Field systems fail in combinations: a camera clock drifts while storage stalls and cellular degrades. The architecture must make those states observable.
Separate work that can stall.
Recording, database writes, uploads, event delivery, retention deletion, analysis, and background services use separate queues or workers so a slow integration does not become a lost video segment.
Expose health as data.
Recording coverage, gaps, stream state, storage pressure, hardware health, service attempts, deployment state, and tunnel status are queryable through the same API surface.
Gate powerful operations.
Destructive API and agent actions use confirmation requirements, filesystem boundaries, SQL restrictions, scopes, audit records, and deployment checks appropriate to the operation.
Cite what supports the answer.
Panoptic answers can carry supporting summary, image, and event identifiers, confidence, and uncertainty. Verification outcomes distinguish support from insufficient evidence.
Do not confuse missing data with no activity.
Where Cognia coverage events are configured, activity, observed-empty time, and unknown/offline periods remain distinct throughout analysis and reporting.
System architecture
A local media plane. A connected control plane.
The exact cloud, sensor, and alert paths are selected per deployment; the edge recording core remains the stable center.
Physical field
Inputs
- RTSP / ONVIF / local cameras
- GPS + light phase
- Weather + LoRaWAN
- Power + cellular
- Digital inputs / outputs
Continuum edge engine
Local truth
- GStreamer capture + media
- DeepStream / TensorRT inference
- SQLite timeline indexes
- Health + retention workers
- REST / WS / MoQ / WHEP surface
Connected operations
Destinations
- Operator applications
- Panoptic search + reporting
- Cloud storage + SFU
- Webhooks + alarms
- Fleet API + MCP agents
Video chooses its own route.
Camera ingest terminates locally. Browser live rides sub-second per-camera streams over Media over QUIC by default, with direct WebRTC, a configured cloud SFU, and HTTP-served media as alternate routes. The archive stays on the node until policy sends it elsewhere.
Operations remain observable.
REST, WebSocket, cloud heartbeat, authenticated remote access, event delivery, deployment controls, and fleet tools connect field state to users and systems.
Workloads do not share one lock.
Configuration, segments, timelapse, events, and logs use separated SQLite databases and dedicated write paths to reduce contention between high-volume retention and operational inserts.
Readings stay local first, too.
LoRaWAN and wired sensor readings are decoded, stored, and threshold-evaluated on the node in their own database lane. Alerts fan out through configured channels; satellite nodes forward to a hub with authenticated, restart-safe delivery.
Security controls
Layered access, constrained integrations, inspectable actions.
SurveillX combines edge credentials, scoped cloud tokens, transport security, signed outbound events, secret handling, and operational audit data. Customer requirements and network placement are reviewed per deployment.
- Per-node API keys and support for scoped, signed cloud authorization
- TLS-protected cloud and remote-access channels
- HMAC-SHA256 signed webhooks and outbound request safeguards
- Secrets refresh, protected configuration handling, and bounded file access
- Request audit, scheduled-task history, deploy history, and service-health telemetry
Powerful tools have boundaries.
Confirmation gates protect deletes and writes. SQL blocks dangerous operations and limits results. File access is sandboxed. Tool metadata and scopes distinguish read from write behavior.
Updates are state machines.
Current deployment flow is digest-addressed, hardware-image compatible, self-tested, historically recorded, and designed with rollback controls—not a blind remote shell command.
Source media retained locallyProtect · validate · export
Evidence posture
Preserve the source. Be exact about the guarantee.
SurveillX provides a strong operational evidence workflow: source segments, indexed time ranges, protection from routine retention, optional hashes, frame extraction, clips, and signed sharing. It does not represent these controls as a certified immutable chain of custody.
- High-resolution local archive with recording-gap visibility
- Segment and timelapse protection from automated retention
- Optional per-segment SHA-256 creation and background validation
- Bounded MP4/TS export and time-limited cloud links
- Sealed, independently verifiable OEFF export with a per-trailer exporter identity—the format NIST recommended as the FBI’s minimum interoperability requirement (NISTIR 8161r1) and IEC standardized (62676-2-32); tamper-evident from the trailer, not a certified chain of custody
AI grounding
Keep retrieval, measurement, and interpretation inspectable.
Connected intelligence is designed to narrow the review and expose its support—not to turn sampled evidence into certainty.
Search within an authorized field.
Customer, site, camera, time, trigger, event type, and detection-class boundaries constrain the evidence considered for an answer.
Use the right representation.
Structured activity supports calculation. Captions and summaries support semantic retrieval. Native image embeddings support visual similarity. Source media remains available for review.
Make uncertainty part of the result.
A claim can be supported, partially supported, not supported, or have insufficient evidence, with confidence and the identifiers used to reach that outcome.
State what the camera could observe.
Activity, observed-empty, and unknown/offline are different facts. Reports carry coverage so a missing window cannot silently become a negative finding.
Keep intelligence records addressable.
Content-addressed event identifiers, evidence references, and stored report results let connected services point back to the same underlying record.
Expose intelligence through the control plane.
Customer-facing search, Ask, image retrieval, report definitions, and run history pass through SurveillX access controls rather than a public Panoptic operator surface.
Capability clarity
What ships broadly—and what depends on the deployment.
This summary reflects the codebase’s maintained capability matrix. A site review confirms exact hardware, credentials, model profile, network path, and rollout state.
| Capability | Availability | Claim boundary |
|---|---|---|
| Edge recording, retention, snapshots, clips | Core | Current platform capability, sized to installed compute and storage. |
| Sub-second live (MoQ), WebRTC, HLS playback, GPU composition | Core | Available; exact live route depends on network and deployment. |
| Person/vehicle detection + ROI rules | Core | Current when the relevant model and stream profile are configured. |
| Vision snapshot analysis + comparison | Core | Current; uses configured provider and may be pay-per-call. |
| Panoptic search, Ask, visual progress, reports | Connected service | Current through authorized, customer-scoped cloud paths; exact workflows depend on account role and connected evidence. |
| Cloud storage, SIA, Immix, Procore photos | Configured | Requires customer credentials and destination-specific setup; Procore scope is photo albums. |
| Weather stations, LoRaWAN, Victron, GPIO | Hardware-specific | Requires compatible field hardware and validation for the target device. |
| Cognia activity learning, rules, anomaly, novelty | Separate stack | Confirm edge deployment, camera policy, model configuration, schedule, and baseline warm-up before demo or quote. |
| Certified evidence immutability / chain of custody | Not claimed | Protection flags and optional file hashes are not a certification. |
| PPE / safety compliance guarantee | Not claimed | Analytics may assist human review; validation and liability controls are required for compliance use. |
| Vehicle-specific SKU and workflows | Roadmap | The edge architecture is reusable; vehicle power, privacy, CAN/OBD, and portal workflows are not a current SKU. |
Technical review
Questions worth asking.
The edge node continues local recording, archive indexing, configured native inference, retention, health monitoring, and LAN-accessible functions. Remote live access, cloud backup, external model calls, and cloud notifications wait on connectivity or follow their configured retry behavior.
Camera ingest and the archive remain local. Real-time live media travels over Media over QUIC by default (with direct WebRTC and cloud-SFU routes as alternates), while API, signaling, snapshots, clips, and other HTTP-served media use managed HTTP paths. Site-originated backups and events go directly to their configured destinations.
Yes. The recommended posture for many enterprise environments is to use SurveillX as the resilient field edge for remote capture, telemetry, recovery, and low-bandwidth access, then connect selected evidence and events into approved enterprise workflows.
Panoptic can return confidence, uncertainty, and supporting summary, event, and image identifiers. Quantitative questions use structured activity data; visual claims use selected evidence and retain sampled-frame limits. Operators can return to Continuum source media for the surrounding time.
No. For industrial and utility environments, the initial boundary is physical situational awareness outside SCADA, protection relays, hazardous process control, and certified safety functions. Any integration into those environments requires separate engineering and buyer security review.
Camera and codec behavior, retention target, power budget, expected backhaul, remote access policy, cloud destinations, alarm paths, exact sensor models, AI model profile, field mounting, and the human response workflow should all be confirmed against the actual site.
Bring your requirements
Review the architecture before the pilot.
We’ll walk through data paths, failure modes, credentials, hardware dependencies, retention, integrations, and claim boundaries with your technical team.