Sentinel Investigation logo
Security Fraud · Edge On Device

Sentinel Investigation

Sentinel Investigation is a forensic facial and object recognition system from SAFEZA AVA-X for investigators, forensic units and security agencies. It searches terabyte-scale video archives to correlate people, vehicles, objects and events, with 99.82% vendor-claimed accuracy on LFW.

Active GDPR compliant Contact Sales No public API Verified by Guidaio
Overview

What is Sentinel Investigation?

Sentinel Investigation is the post-event video analysis product in the Sentinel Suite from SAFEZA AVA-X. Where Sentinel Live watches streams as they happen and Sentinel Access handles biometric entry at doors and turnstiles, this one is built for what comes after an incident: searching, matching and correlating footage that has already been recorded. The wider range also includes Analyzer, Heliospan and Keystone Identity.

The job is deliberately narrow. Investigators and analysts feed it large volumes of video and images, then ask it to correlate persons, vehicles, objects and events across multiple sources. The vendor describes a four-stage pipeline: Connect Data Sources, AI-Driven Face Matching, Instant Identity Verification and Actionable Forensic Data. Underneath, SAFEZA AVA-X cites parallel GPU processing, in-house indexing and compression, and fine-tuned neural networks.

The performance figures all come from the vendor, with no accessible independent test behind them. SAFEZA AVA-X claims 99.82% accuracy on the LFW benchmark against 99.63% for Google FaceNet and a 97.50% industry average in its own comparison table, plus query latency under 0.5 seconds, 10M+ searchable facial images and terabyte-scale archives. A vendor blog post puts the practical gain at work taking analysts 48 hours completed in under 15 minutes. Read all of these as claims, not measurements.

The deployment model is what really distinguishes it. The software runs on GPU servers the customer owns, and the vendor states it can operate fully air-gapped, with model weights stored on those same customer-controlled machines. It connects to existing IP cameras over RTSP, ONVIF or a stream URL and hooks into a VMS already in place, rather than replacing the camera fleet. The product page advertises cloud and on-premises deployment, yet the site menu still marks "Sentinel Cloud" as Coming soon.

The vendor positions the tool as operator-in-the-loop: matches and alerts are presented for human verification before anyone acts on them. It is designed and developed in Switzerland and, according to SAFEZA AVA-X, trained and proven on South African data. Nothing about it is self-service: no published price, no advertised free trial and no publicly available API documentation.

What it does

  • Ingest footage from local drives, network storage or a VMS in standard media formats
  • Search for a reference face across millions of images using deep-learning embeddings
  • Correlate persons, vehicles, objects and events across several sources at once
  • Read confidence scores and visual match previews built for human verification
  • Export detailed forensic reports and share them between departments
  • Build reusable watchlists for operations that are still running
  • Work on degraded footage: low light, awkward angles, partial occlusion and motion blur
Audience

When to use Sentinel Investigation / When not to

A quick filter to help you decide if Sentinel Investigation is the right fit.

When to use Sentinel Investigation

  • Criminal investigation teams that need to find one face across weeks of multi-camera CCTV footage
  • Forensic units that must produce exportable, chain-of-custody evidence for court proceedings
  • Organisations with strict data sovereignty requirements: processing runs on infrastructure the customer owns, including fully air-gapped sites
  • Sites that already run a mixed fleet of IP cameras (RTSP/ONVIF) and an existing VMS they want to keep
  • Critical infrastructure operators: ports, airports, mines, public transport networks and stadiums

When not to use Sentinel Investigation

  • Teams that need real-time monitoring of live streams: that is Sentinel Live, a separate product in the same suite
  • Buyers looking for biometric access control at doors and turnstiles: that is Sentinel Access, not this product
  • Analysts after text or OSINT analysis: that job belongs to Heliospan, another product in the range
  • Anyone expecting a published price or a self-service sign-up: nothing is listed and every deal goes through a quote
  • Small organisations without GPU servers or a security team: inference runs on hardware the customer owns and operates
Get started

How to use Sentinel Investigation

A typical end-to-end flow, from setup to results.

  1. Start on the pricing page: a configurator asks for the product, the AI models, the number of cameras and the deployment model, and opens with your email address
  2. Discuss a scoped pilot with the vendor, typically on one defined set of cameras and one workflow at a real site such as a neighbourhood, campus, mine gate or terminal
  3. Go through the architecture review: camera streams, VMS hook-in points and a first watchlist
  4. Size the hardware: GPU servers you own, dimensioned by camera count, retention and the mix of products deployed
  5. Connect the data sources: local drives, network storage, IP cameras over RTSP or ONVIF, and your existing VMS
  6. Load the reference faces and build the watchlists the investigation needs
  7. Run searches across the archive and review the confidence scores and visual previews returned
  8. Have an operator verify every match before any action is taken on it
  9. Export the forensic report, with chain of custody, and share it with the departments that need it
  10. Keep the deployment current with cryptographically signed update packages, installable offline at sites without permanent connectivity
Quick read

Pros & Cons

Pros

  • Real data sovereignty: processing and model weights sit on customer-controlled infrastructure, and the vendor states the system can run fully air-gapped
  • No raw video retention by the vendor: footage stays in the customer's VMS under the customer's own retention policy
  • No permanent vendor backdoor: remote access is opt-in, time-limited and governed by the customer's rules
  • Broad compatibility with existing kit: IP cameras over RTSP/ONVIF, a VMS already in place, and 31 camera brands listed as compatible
  • Court-oriented traceability: evidence export with chain of custody, plus audit trails on detections
  • Claimed operator-in-the-loop design, backed by a public page setting out AI ethics and GDPR/POPIA commitments
  • A published, numbered accuracy benchmark with a named point of comparison (Google FaceNet)

Cons

  • No public pricing at all: the pricing page states outright that no list prices are published
  • No publicly available API documentation, even though "Seamless API Integration" is advertised as a key feature
  • No separate contractual documents: the three footer links (Privacy Policy, Consent Declaration, Terms of Service) all resolve to the same POPIA privacy policy
  • The site contradicts itself on cloud: "Sentinel Cloud" and the Service Desk are marked Coming soon while the product page advertises cloud deployment
  • The performance figures (99.82%, under 0.5 s, 10M+ images) are published by the vendor with no independent test report available
  • Heavy hardware prerequisites: GPU servers owned by the customer and sized to the camera fleet
  • No minimum age, no customer DPA and no named list of sub-processors published anywhere on the site
Pricing

Pricing & Plans

No free plan and no free trial are advertised, and no entry price exists to quote: the pricing page states in plain words that no list prices are published, and its meta description repeats "No list prices published". Pricing is entirely quote-driven. The site offers a configurator where a prospect selects the SAFEZA AVA-X product, the AI models, the camera package and the deployment model, then submits contact details, after which the vendor undertakes to return commercial terms for that specific deployment. A scoped pilot is offered as an entry point, but it is nowhere described as free of charge. African contracting and first-line support are handled by SAFEZA AVA-X, the South African entity. Any figure will therefore come out of a sales conversation, not from the website.

Prices and plans listed above may evolve. Always check the official pricing page before subscribing.
Trust & Privacy

Data, GDPR & hosting

A consolidated view of how Sentinel Investigation handles your data.

GDPR overview

The reference document here is POPIA, not the GDPR. SAFEZA AVA-X (PTY) LTD publishes a South African privacy policy and says it also takes account of foreign data protection laws, including the EU GDPR and Swiss law, only to the extent they apply to it, which makes the GDPR secondary rather than governing. A dedicated "GDPR, POPIA & AI Ethics" page claims alignment with both regimes from design through deployment and sets out six principles: Fairness, Transparency, Accountability, Non-Discrimination, Notice and Consent, and Lawful Surveillance. The vendor also claims regular third-party algorithm audits and demographic bias testing. No EU representative under Article 27 is named anywhere, and no DPO is appointed: the published contact is the POPIA Information Officer at privacy@safeza.ai. Data subject rights are described in POPIA terms, with the Information Regulator of South Africa as the escalation route.

Who owns the data?

Under the privacy policy issued by SAFEZA AVA-X (PTY) LTD, the customer deploying the technology is normally the responsible party for personal information processed in its own environment; the vendor may act only as an operator, a processor working on written instructions, in line with POPIA. Operationally the split matches: video, biometric templates, watchlists and operational records stay on the customer's infrastructure, so the customer keeps custody as well as responsibility. At the end of a contract the data remains with the customer, while the vendor revokes its access and removes its software. The published policy is governed by South African law: POPIA primarily, with PAIA and ECTA also cited.

Reuse rights

The stated purposes are POPIA-based: performing the contract, meeting legal obligations, legitimate interests, and consent for marketing and non-essential trackers. The website itself collects identity and contact details, contractual and billing documents, technical data such as IP address, browser, operating system and logs, plus information drawn from public registers. Biometric information is treated as special personal information under POPIA sections 26 to 33, and transfers outside South Africa are framed by section 72. On the product side, the AI engine is described as trained on millions of diverse, ethically sourced facial datasets belonging to the vendor. What the site never states is whether a customer's own footage may be used to train or refine models: no position is published either way, and no documented opt-out exists, so this is a point to settle contractually rather than assume.

Data retention & training

Retention summary
Two separate regimes apply. For data the website collects, the POPIA privacy policy says personal information is kept only as long as needed for the purpose it was collected for, or as long as South African law requires, including tax and company record obligations. After that it is destroyed, deleted or de-identified in a way that prevents reconstruction. Where one record serves several purposes, access is restricted to the purpose still active until every retention period has expired. For operational data the picture is different: the vendor keeps no raw video, so retention of recordings stays inside the customer's VMS and under the customer's own policy. When a contract ends, the data stays with the customer while the vendor revokes its access and removes its software. No retention period is published in days or months.
Trains on customer data
Unclear
GDPR contact

Hosting summary

There is no vendor hosting to describe. SAFEZA AVA-X states that processing runs on infrastructure the customer controls, in one of three forms: GPU servers on the customer's own premises, a designated in-country installation, or a regional cloud data centre the customer has approved. Model weights are installed on those same customer-controlled servers, typically inside the customer's data centre, control room or approved regional hosting. Telemetry can be switched off; where it is enabled, the vendor describes it as scoped and auditable, with nothing leaving the host country without authorisation. Encryption in transit and at rest, role-based access control and MFA where applicable are applied across that customer-controlled estate. The vendor explicitly says it does not deliver a multi-tenant offshore SaaS for sovereign installations. That is why no vendor hosting country or region is recorded on this page: the jurisdiction that applies is whichever one the customer picks for its own servers. Buyers who need a guarantee on data residency should have it written into the deployment agreement rather than infer it from the website.

Watch-outs

Things to keep in mind

Risks and trade-offs to weigh before adopting Sentinel Investigation.

  • This is biometric facial recognition, and the vendor's own ethics page acknowledges the risk of errors, bias and misuse rather than denying it
  • Biometric information is special personal information under POPIA: the deploying customer stays responsible for its own compliance, including notification, justification and any required authorisations
  • Human oversight is a duty the customer carries, not something the software discharges; the site is explicit that lawful use and supervision remain the customer's responsibility
  • Confidence thresholds are a genuine hazard: the vendor itself publishes an article about a wrongful arrest based on a 93% match
  • Automation bias is easy to slip into: a ranked list of matches invites confirmation rather than investigation, which is precisely what the operator-in-the-loop design is meant to counter
  • No minimum age is published, and the site says nothing about how minors captured on camera are handled
  • Check the legal framework that applies locally before deploying any facial recognition system: the rules differ sharply between jurisdictions
Setup

Setup & Integrations

Technical difficulty

High. This is an infrastructure project, not a sign-up, with no self-service registration. You need GPU servers you own, sized to camera count, retention and product mix, plus an architecture review covering stream sources, VMS hook-in points, a first watchlist and cipher suites aligned to your standards. Existing IP cameras connect over RTSP, ONVIF or a stream URL, so the fleet stays. The vendor factors in field conditions: power quality, generators, rack space, intermittent connectivity. Expect to begin with a scoped pilot and to involve security, legal, ICT and procurement teams.

Integrations

NVIDIA Jetson HPE Telegram Threema Kafka
Company

Behind Sentinel Investigation

Company name
SAFEZA AVA-X (PTY) LTD
Founded
INFORMATION_NOT_FOUND
Country of origin
🇿🇦 South Africa
Headquarters
Kensington, Cape Town, 7405, South Africa
EU office
Bockenheimer Straße, Frankfurt am Main, Germany
US office
Delaware, United States
UBO
INFORMATION_NOT_FOUND
UBO country
INFORMATION_NOT_FOUND
Domain registrar country
🇺🇸 United States

Social

Official links

Resources

All the official URLs gathered for verification and reference.

Compare

Alternatives

Tools that compete with or complement Sentinel Investigation.

G Google FaceNet
FAQ

Frequently asked questions

Where are the video and biometric data processed?
On infrastructure the customer controls: GPU servers on the customer's own site, a designated in-country installation, or a regional data centre the customer has approved. The vendor states there is no obligation to send data to a foreign vendor cloud.
Does SAFEZA AVA-X keep the raw video?
No. According to the vendor, the tool produces alerts, matches and metadata; retention of the recordings themselves stays inside the customer's VMS and under the customer's own policy.
Can Sentinel Investigation run completely offline, in an air-gapped environment?
Yes, according to the vendor: processing is local, model weights are stored locally, and updates are installed offline through cryptographically signed packages carried on removable media.
Does it work with the cameras and the VMS already installed?
Yes. It connects over RTSP, ONVIF or a stream URL, and the vendor publishes a list of 31 camera brands described as compatible. That page is explicit that presence on it means typical network configuration, not an exclusive partnership.
What is the difference between Sentinel Live, Sentinel Investigation and Sentinel Access?
Sentinel Live monitors streams in real time, Sentinel Investigation searches and correlates archived video after an incident, and Sentinel Access handles biometrics at doors and turnstiles. They are three separate products in the same suite.
Can an operator override a decision made by the AI?
Yes. The vendor describes the tool as operator-in-the-loop: alerts and matches are presented for human verification, and a person is expected to confirm before any action is taken.
Can detections be audited for an investigation or a court case?
The vendor states that searches and correlations can be audited and that evidence can be exported with a chain of custody, which is the point of the forensic output stage of the pipeline.
Who can access a deployment remotely?
Only people the customer authorises, according to the vendor. Supplier remote access is described as opt-in and time-limited, with no permanent vendor backdoor.
What hardware is required on site?
GPU servers owned by the customer, sized to the number of cameras, the retention required and the mix of products deployed. This is a server-side installation, not a desktop or mobile application.
How much does Sentinel Investigation cost?
No list price is published. Pricing is obtained by submitting a configuration through the pricing page, after which the vendor returns commercial terms for that specific deployment.
Conclusion

Should you pick Sentinel Investigation?

Sentinel Investigation is a narrow, vertical tool rather than a general-purpose product: it exists to run forensic searches across recorded video, and it deserves to be judged on that job alone. Its structural strength is the deployment model. Processing, model weights and biometric templates stay on infrastructure the customer owns, the system can reportedly run air-gapped, the vendor keeps no raw video, and supplier remote access is opt-in and time-limited. For a police force, a prosecutor's office or a critical infrastructure operator that cannot send footage to a foreign cloud, that architecture is the entire argument, and it is unusually well documented in the vendor's own FAQ.

Its structural weakness is commercial opacity. No price is published anywhere. There is no separate terms of service: all three footer links resolve to the same POPIA privacy policy. There is no publicly available API documentation, although API integration is advertised as a key feature. The cloud option promoted on the product page is marked Coming soon elsewhere on the site. Each of these gaps can be closed in a sales conversation, but none can be checked before you start one.

The numbers deserve the same caution. 99.82% on LFW, queries under half a second, 10M+ searchable faces, 48 hours of analyst work compressed into 15 minutes: every one of these is published by SAFEZA AVA-X, and none is backed by an accessible independent test report. Ask for one during the pilot.

The realistic buyer is institutional, with security, legal, ICT and procurement teams around the table and GPU servers in the budget. Anyone hoping to sign up online, compare tiers and try it out will not find that here.