Jacek Janczura Founder-CTO · Enterprise AI
Book a call

Case study · Enterprise AI platform

entAIngine: one multi-tenant AI platform behind every client deployment

The platform every client deployment runs on: multi-tenant, event-driven on AWS, with a provider-independent model layer, role-based access control, a retrieval service over client documents, and its own tooling for prompt optimisation and pipeline evaluation.

Production AI for your workflow All case studies

Own company Production

Mandate

Architected the platform and led its implementation: the multi-tenant model, the event-driven runtime, the provider-independent model layer, access control, the retrieval service and the tooling around prompts and evaluation.

Key decisions

  1. Keep the model layer provider-independent (AWS Bedrock and SageMaker, Azure, Google Gemini, OpenAI, Mistral) so model choice follows quality, latency, cost and data sovereignty per client.
  2. Event-driven serverless architecture on AWS rather than long-running services, so cost tracks usage and the platform autoscales into peaks.
  3. Build evaluation and prompt tooling as platform primitives (Testbed, Prompt Wizard) instead of letting every project write its own scripts.

Outcome — measured

50+ organisations running AI processes on one multi-tenant platform
99.9% platform uptime
>1,000 concurrent connections and thousands of requests per second
ISO 27001 + SOC 2 Type 2 certified, with data sovereignty for regulated clients
Stakeholders
  • Enterprise clients in manufacturing, pharma, financial services and back-office operations
  • Client IT, security and compliance teams
  • Frontend, backend and AI workstreams inside the company
  • External auditors
Constraints
  • Multi-tenant isolation with data sovereignty per client
  • ISO 27001 and SOC 2 Type 2 controls
  • Model providers changing capability and price faster than release cycles
  • Cost efficiency under bursty enterprise load
Reuse
  • Provider-independent model layer
  • Retrieval service that builds vector stores from client files
  • Testbed: deterministic test cases for AI processes and retrieval pipelines
  • Prompt Wizard: automated prompt optimisation
  • MCP-based tools exposing platform capabilities to agents and clients

Context

I was co-founder & CTO of Jaden Data until June 2024 and have been its fractional CTO and consulting partner since. entAIngine is the SaaS platform for AI process automation that Jaden Data built and that every client deployment in these case studies runs on: knowledge-aware assistants, document workflows and back-office automation for organisations in manufacturing, pharma and financial services.

Business problem

Each enterprise arrived with its own security review, its own document estate and its own preferred cloud. Solving that per client produces bespoke stacks that never pay for themselves. The platform had to make a new deployment mostly a matter of configuration, while still meeting the security and data-sovereignty bar that regulated buyers set.

My mandate

I architected the platform and led its implementation: the multi-tenant model, the event-driven runtime, the model integration layer, role-based access control, the retrieval service, and the prompt and evaluation tooling around them.

Decisions

  • A provider-independent model layer integrating AWS Bedrock and SageMaker, Azure ML Studio and AI Foundry, Google Gemini, OpenAI and Mistral, so a client’s model choice follows quality, latency, cost and data-residency rules rather than a vendor commitment made once.
  • Event-driven microservices on AWS — Fargate on ECS with ECR, Lambda, SNS and SQS, CloudWatch, EC2 and S3 — audited against the AWS Well-Architected Framework and tuned for serverless autoscaling.
  • Role-based access control and per-tenant boundaries as platform primitives rather than per-project code, because every enterprise security review asks the same questions.
  • A retrieval service that builds vector databases from client files automatically, with Pinecone, pgvector and Milvus behind one interface, so document retrieval was a configuration choice.
  • Testbed, a testing framework that assesses AI processes and retrieval pipelines with deterministic test cases, quantitatively and qualitatively — so a prompt change could be judged, not argued about.
  • Prompt Wizard, which automates the optimisation of prompts, so prompt engineering did not depend on one person’s intuition.
  • MCP-based tools deployed on the platform, making its capabilities available to MCP-compatible agents and clients.

Delivery

The platform ran client deployments in insurance, legal, automotive and property alongside the work documented in the other case studies. It reached more than 1,000 concurrent connections and thousands of requests per second at 99.9% uptime, under ISO 27001 and SOC 2 Type 2.

Outcome

One platform serving 50+ organisations, with the security posture regulated buyers require and a cost profile that follows usage rather than provisioning.

Reuse

Everything in the platform exists because a client project needed it twice. The model layer, the retrieval service, Testbed, Prompt Wizard and the MCP tools were built once and used in every later deployment.

Evidence

Organisation count, uptime, concurrency and throughput are company records; the certifications are audited. The client-side results the platform produced are documented in the individual case studies.

Want the same thing done in your environment?

This case is one of several. If the shape looks like your problem, the fastest route is to send me the constraints you cannot move.

Send a project brief All case studies

Remote-first, on-site when it matters; NDA on request