Cloud architecture, backend systems and GenAI integration that your team can build on.

Challenge

Industrial customers model their plants as typed entities, connect live machine data over MQTT, IoT Hub or file drops, bind signals to assets and run analytics flows from a node-based flow editor. The platform had to be rebuilt as a product that runs both as SaaS and as a bring-your-own-cloud install on customer infrastructure, with a real identity model, regional data residency and infrastructure that can be reproduced from code.

Approach

• Set up the new monorepo and owned architecture, infrastructure and core backend components as external architect, hands-on as the most active contributor.
• Tenancy: one database per tenant with a tenant registry in SaaS mode, single-tenant Docker Compose install for BYOC, both served by a shared tenancy library.
• Identity on self-hosted Zitadel with OIDC: one organisation per tenant with admin, member and viewer roles, branded login per tenant, idempotent tenant onboarding through a dedicated user management service, internal tokens between services.
• Infrastructure as code with Terraform: one stack module per environment on Azure Container Apps, Service Bus, IoT Hub and Key Vault, private network with a single public proxy, secrets generated on first apply and never stored in variable files.
• Data moved from Cosmos DB to MongoDB Atlas with regional clusters for data residency and native time series; Prefect 3 as workflow engine triggered by internal events.
• Split ingest out of the flow engine into its own service with one adapter per source and certificate-based MQTT ingress with per-tenant client certificates. Frontend moved from Angular to React.

Outcome

• SaaS and BYOC run from one codebase, tenant and developer onboarding are automated.
• Network locked down to one public entry point with a fixed egress IP, Terraform drift ended, security hardening across proxy, API docs, secrets and service permissions.
• Ingest path load-tested for more than ten thousand data points per second.
• Architecture papers, decision records, reviews and mentoring, plus shared Azure DevOps pipeline templates with per-package releases, so the team can carry the platform forward.

Challenge

The product was being rebuilt from a collaborative ideation tool on Angular, Node.js and MongoDB into an AI video platform: users write a script, get a synopsis and a storyboard and generate AI images and videos per shot. The legacy user management could not carry subscriptions, credits or a login shared across the new services.

Approach

• Designed and built a dedicated user management service with FastAPI and PostgreSQL as the central hub for authentication, users, subscriptions, credits and billing. The video backend only verifies sessions.
• Authentication on SuperTokens with email and password plus Google sign-in, HttpOnly session cookies shared across the platform's subdomains, email verification and password reset.
• Stripe checkout, customer portal and webhooks. Credit model with subscription credits per billing period and purchased credits with twelve-month expiry, reserved on request and confirmed or refunded afterwards, because a video generation can run for more than twenty minutes.
• Built the React storyboard and video editor: scene and shot playback, reordering, trimming, versioned variants, image-to-video, PDF and ZIP export.
• Structured JSON logging with redaction of sensitive data, Alembic migrations for both Python services, test suites for credits, payments, plans and admin functions.

Outcome

• Legacy user management removed, subscriptions and credits live on Stripe.
• New React frontend with a feature-based architecture that the team keeps extending.
• Billing and credit model documented with diagrams so product and engineering share one picture.