Building the Predictive Maintenance AI Platform That an Industrial IoT Startup Took to Series A

Streaming ingestion and anomaly detection: the foundation that fires before alarms do

The first architectural decision set the constraint for everything else: anomalies had to be surfaced from raw sensor data before threshold alarms or human observation would catch them.

Ideas2IT built

1. the ingestion layer on Azure IoT Hub and Azure Stream Analytics, receiving terabytes of sensor data in real time, with Kepware abstracting plant equipment data into the hub.
2. The anomaly detection layer ran Random Forest models with variational autoencoder architecture in Azure ML Studio on Azure AKS.
3. A Celery-based rule engine triggered contextual alerts based on business-defined thresholds and routed notifications to operators via web, Teams, and email.

This Phase Produced

• Azure IoT Hub ingestion pipeline (Real-time sensor data ingestion from manufacturing plant equipment)
• Azure Stream Analytics resampling layer (Parallel computation prep for ML processing)
• ML anomaly detection models (Random Forest + variational autoencoder, Azure ML Studio on AKS)
• Celery-based rule engine (Business-condition alert triggering with configurable thresholds)
• Operator notification system (Web, Microsoft Teams, and Sendgrid email routing)
• End-to-end Azure infrastructure (IoT Hub, Event Hub, Stream Analytics, Azure Functions, Timescale DB)

Multi-tenant architecture and microservices layer: 300% infra cost reduction at scale

The platform needed to serve multiple manufacturing customers without duplicating infrastructure per tenant.

Ideas2IT designed

1. a cost-effective multi-tenant architecture with isolated Timescale DB instances per tenant, backed by Azure Storage for durability.
2. Ten-plus OpenAPI FastAPI microservices handled data ingestion, retrieval, analytics, visualization, and operator notifications.
3. The team migrated from Flask to FastAPI when the streaming workload demanded it, reducing CPU and memory consumption measurably.
4. The multi-tenant model combined with a redesigned Azure ML Studio deployment reduced overall infrastructure costs 300% against the single-tenant baseline. Timescale DB holding 1.5TB of historical data was tuned to return content in under 300ms.

This Phase Produced

• Multi-tenant Timescale DB architecture (Per-tenant database isolation with Azure Storage backup)
• 10+ OpenAPI FastAPI microservices (Ingestion, retrieval, analytics, visualization, notification)
• Flask-to-FastAPI migration (CPU and memory reduction under streaming workload)
• Azure ML Studio deployment redesign (70% reduction in ML infrastructure costs)
• Timescale DB query optimization (Sub-300ms load for 1.5TB historical dataset)
• Admin Manager UI (Platform-level tenant and user management interface)

Security architecture and CI/CD hardening: the compliance foundation for enterprise customers

Manufacturing customers require compliance postures that survive platform upgrades and hold up under external audit. Ideas2IT implemented Istio request authentication and mutual TLS, isolating the authentication layer from regular stack changes so an upgrade could not introduce a security regression.

Auth0 with JWT handled application-level authentication. A half-yearly third-party security audit cycle runs with findings addressed inside the subsequent sprint. Terraform and Helm charts codified the full Azure infrastructure as reproducible manifests.

End-to-end tests ran via Playwright with automated regression capture, and streaming logic changes were validated through a custom streaming simulator before reaching production.

This Phase Produced

• Istio mTLS + request authentication (Auth isolated from upgrade/stack change cycles)
• Auth0 + JWT application authentication (Application-level access control)
• Half-yearly third-party security audit (Findings addressed in-sprint)
• Terraform + Helm infrastructure manifests (Full Azure infra as reproducible code)
• Playwright end-to-end test suite (Automated regression capture on each release)
• Custom streaming simulator (Validates streaming logic changes before production)