Project · SaaS · Live in production
AICloud
The cloud control room that watches every AIBox in the field - alerts, evidence, escalations, and a live map of every site, all in one dashboard.
- Period
- Jul 2025 - Jul 2026
- Role
- Architect and lead full-stack engineer
- Client
- PathTech JSC - serving Petrolimex Aviation. Live at aivision.petrolimexaviation.com.
10
backend services in production
15+
operator surfaces in the web dashboard
2
client surfaces (web + mobile)
1-command
fleet deploy across 3 nodes
Problem
Edge AI boxes at airports and substations spot real problems all day - missing helmets, oil spills, smoke, unauthorized people - but until something reaches a human, nothing happens. Operations needed a single cloud surface to watch every site at once, see the evidence photo for every alert, route serious incidents to the right person within an SLA, and let admins manage cameras, AI models, roles, and licenses without ever SSHing into a device.
Approach
- 01
Designed a microservices backend split by concern: identity (login, roles, permissions), devices (the AIBox fleet), events (every detection), incidents (grouping events into actionable cases with SLA timers), notifications, licensing, organizations, and a real-time push service that streams alerts to browsers within a second.
- 02
Built the operator surface in React with a live map, a filterable alerts feed with the AI-annotated photo and short video clip per event, and a tri-state assessment workflow (Unassessed → True / False) so the team's review actions can feed retraining later.
- 03
Designed a four-layer permission model - Branch → AIBox → Camera → AI Model - so a regional supervisor only sees their own region and a contractor only sees their own camera. Roles cannot grant rights they don't already hold.
- 04
Shipped a companion mobile app for on-call operators, with 90-day sessions for in-the-field reliability and a join-request flow for new accounts.
- 05
Wrote a single deploy script that pulls the latest code, rebuilds every service image in parallel, runs database migrations safely, deploys to a three-machine cluster, and verifies health - turns a release from a half-day of work into one command.
From the live product

Dashboard - every site at a glance.

Cameras - latest AI-annotated frame per stream.

AI models - the catalog of what's running where.

AIBox onboarding - new edge devices come in here.
Stack
- NestJS · TypeScript · pnpm workspaces
- PostgreSQL · MikroORM · Redis
- Keycloak (OIDC auth)
- NATS (service-to-service) · EMQX (MQTT to the edge)
- WebSocket real-time push
- React 19 · Vite · TanStack Router/Query · oRPC contracts
- React Native mobile app
- Docker Swarm · Kong gateway
Lessons
- Multi-tenant isolation is a discipline, not a feature. Every business request carries an organization context and admins never see anything outside it - enforced at the API layer, the UI layer, and the role editor.
- Real-time over WebSocket is the right abstraction for a 'control room' UI. Polling for alerts is dead the moment a customer measures it in seconds.
- Operators don't want one more alert - they want fewer, with better context. Grouping events into incidents with SLA timers is what made the feed actually usable in the field.