The vision AI platform for enterprise.
Build, deploy and monetise computer-vision solutions across any domain — on the cameras you already own. Describe what to watch for in plain English, prove it on your own recorded footage, then run it.
Every vision AI use case is rebuilt from scratch. That is why vision AI has not scaled.
India has a very large installed camera base and no platform layer on top of it. The missing piece is not another application. It is the thing other applications get built on.
Every use case is a project
Attendance, PPE, traffic violation, classroom compliance — each means a separate vendor, model, integration and a three-to-six-month build.
Single-purpose products
An attendance vendor cannot do safety. A VMS cannot do attendance. One site ends up running four or five disconnected systems.
No self-serve
Changing what the system watches for requires the vendor’s engineers. Customers are locked in and slow; vendors cannot scale past their delivery capacity.
Domain expertise is stranded
The integrator who understands traffic enforcement or school safety has no way to productise that knowledge. The people closest to the problem cannot build for it.
One platform. Any vision AI domain. Built by us, by customers, and by partners.
Five layers. Each capability is marked for what it does today, not what it will do. The two layers with a rule beside them are where the proprietary engineering sits.
Describe what to watch for in plain English. The platform validates it, proves it on your own recorded footage, then runs it.
Build and sell your own branded vision solutions on our infrastructure. B2B2C without building a vision stack.
Publish what you build to the marketplace and earn from it when another tenant adopts it.
Customer model training and live rule execution are next phase. Authoring, validation and simulation are live and validated on staging as of August 2026.
A new vision AI solution in hours. Configuration, not engineering.
The traditional approach is three to six months and a vendor engineering project per use case. Here the customer authors it, verifies it against real footage, and deploys it.
Describe
An operator types, in plain English: alert me when a person is in the loading bay while a forklift is moving, for more than 30 seconds, without a hard hat.
Author
The platform’s agent drafts a structured activity rule and explains it back in plain language.
Verify
The rule is validated, checked for feasibility against the models actually deployed on that site, and simulated over the customer’s own recorded footage.
Deploy
The operator confirms, the rule is bound to cameras and areas, and it is enabled.
Publish
The finished template, optionally with a tuned model, is published for other tenants to adopt, with revenue share to its author.
┌───┐
│ │ │ cortexvigil · activity studio
└─┘ simulate before enable
RULE [loading-bay-ppe]
person in [BAY-2] while forklift moving
> 30s without hard hat
FEASIBILITY
person detection deployed on this site ok
forklift class deployed on this site ok
hard-hat attribute deployed on this site ok
camera [CAM-04] ppm 38 px/m at [BAY-2] ok
camera [CAM-09] ppm 11 px/m — below task declined
SIMULATE replayed [6 h] of your own footage
would have fired [3] times
[09:14:22] 2 people, 41s snapshot
[11:02:07] 1 person, 33s snapshot
[15:48:51] 1 person, 67s snapshot
enable on [CAM-04]? the platform will not enable
[CAM-09]: its models can never fire this rule. The platform refuses to enable a rule its own models can never fire. That trust gate is what makes self-serve authoring safe to put in a non-expert’s hands.
Without it, plain-English authoring produces rules that fail silently in production and nobody finds out for a month.
The authoring language is a bounded whitelist AST. Generated output can never become arbitrary code execution — which is what makes agent-authored rules safe on a multi-tenant platform.
03 — PROPRIETARY PLATFORM ENGINEERING
Not a wrapper around an off-the-shelf model.
Each component draws on established machine-learning technique. The integrated platform, and its application to multi-tenant multi-domain vision, is our own.
Safe rule-authoring language
A bounded whitelist AST. Plain-English input becomes detection logic that can never become arbitrary code execution. The enabling innovation for the whole platform model.
Simulate before enable
Every rule replayed against real footage on the same engine that will run it live, cross-checked against the models actually deployed on that site.
PPM-aware detection confidence
Pixels-per-metre computed from resolution, field of view and mounting geometry; thresholds adjusted per zone; operators warned when a camera physically cannot support the task.
Hierarchical multi-model inference
Model tiers held in a registry and selected in real time against available GPU, scene complexity and task need. Hot-swapped without interrupting the stream.
Cross-camera identity resolution
The same person tracked across multiple cameras and sites under strict per-tenant isolation, where naive per-camera tracking assigns unrelated identities.
Hybrid edge and cloud parity
The identical stack on an on-premise GPU appliance or in our managed cloud. Data-sovereign deployment with no separate product and no re-engineering.
Operator-in-the-loop adaptation
Operator corrections feed an active-learning pipeline with shadow deployment and automatic rollback on regression. Accuracy improves per site over time.
Native agentic layer
137 tools across 27 domains on a production Model Context Protocol server. The platform is authorable and queryable by external AI agents, so customers drive it from their own tooling.
Proprietary. Nothing here is filed; no patent is claimed.
Platform built and validated. First domain live in a production pilot.
Everything on this list exists and runs. What is not on it is on the roadmap page, marked as such.
Self-serve Activity Studio with conversational authoring.
A nine-template catalogue customers instantiate into working solutions.
The full author, validate, feasibility-check, simulate and deploy loop.
Rule versioning, history and restore, with per-tenant isolation.
137 agent-callable tools across 27 domains on the production MCP server.
Five heavy-machinery classes with no customer training required — excavator, dump truck, wheel loader, backhoe loader, bulldozer.
Production multi-tenant cloud and on-premise edge deployments, both live.
Every shipped rule asserted in CI to be reachable and feasible before it can be enabled.
The marketplace — cross-tenant publishing of activity templates and tuned models, with revenue share.
Customer-facing model training and fine-tuning.
The live rule execution engine.
A platform that will not tell you where the line is has already told you something. Marking the roadmap as roadmap costs nothing and is the only version of this list worth reading.
Bring one site and one thing you want watched.
We will author it with you in the Studio, simulate it against your own footage, and show you what it would have caught last week. On the cameras you already own.