A senior engineer's knowledge, on call for everyone.
A multimodal co-pilot inside the client's engineering software that answers engineers from more than 10,000 verified documents, diagrams included.
- Grounded: every answer comes from the client's own manuals, schematics and SOPs, with sources.
- Multimodal: it reads circuit diagrams and drawings, not only the words around them.
- Governed: identity, safety and the agent lifecycle run in Microsoft AI Foundry.
Meet our client
- Client
- Fortune 500 industrial automation company
- Industry
- Industrial automation & engineering technology
- Market
- Fortune 500 enterprise
In a nutshell
We built a multimodal retrieval-augmented generation (RAG) co-pilot for a Fortune 500 industrial automation company's engineering software. It answers automation engineers' technical questions from a verified knowledge base of more than 10,000 specifications, compliance guidelines, manuals, CAD schematics and SOPs, using Google Cloud for retrieval and Microsoft AI Foundry to govern every answer.
All roads led to the same few engineers
Knowledge sat in a handful of senior heads. Every question queued at their desks. Switch on the co-pilot and watch the queue change.
Up to 40% of senior time went on repetitive questions and checking other people's configuration logic.
- A knowledge bottleneck: when senior engineers were busy, away or left, their knowledge became hard to reach and work slowed.
- Repeat questions: up to 40% of senior engineers' time went on repetitive questions and logic checks.
- A steep learning curve: junior engineers needed close oversight to avoid configuration errors.
- Slow change: modifying a control topology or adding a production line meant weeks of manual code checks, with configuration logic reviewed by hand for errors.
Not a chatbot. A first-line engineer.
Three objectives shaped the build, and six design principles kept it industrial-grade. Tap an objective to see what had to change.
Google Cloud retrieves. Microsoft AI Foundry governs.
Two paths. The build path turns documents into a searchable index once. The answer path runs every question through governance. Pick a path to trace it.
- The build path: Cloud Storage to Vertex AI RAG Engine to RagManaged DB through native connectors, built once and kept in sync.
- The answer path: the engineering software to AI Foundry to a foundation model, with retrieval over a secure API gateway.
- Foundry at the centre: every request runs under the engineer's Entra ID identity and passes guardrails on the way in and out.
- Open on models: Gemini or GPT can write the answer; the governance around it stays the same.
From 10,000 files to a searchable index
Steps 1 to 4 happen once and stay in sync. Step through them to see a schematic become something the co-pilot can reason over.
- One verified source: manuals, CAD schematics, SOPs, specifications and compliance guidelines in Cloud Storage, synced from Google Drive.
- Pictures carry answers: the Gemini multimodal parser turns charts, circuit topologies and drawings into descriptive text and vectors.
- Configured, not coded: Vertex AI Studio handles chunking and embeddings; Cloud Spanner scales with built-in sharding and strict isolation.
- Behaviour outside the code: prompts, guardrails and output rules are versioned in the console and change without redeploying.
From question to governed answer in seconds
Steps 5 to 7 run every time an engineer asks. Pick a request and replay it through every service it touches.
- Step 5, the engineer asks: inside the engineering software; Foundry checks identity through Entra ID and applies safety controls first.
- Step 6, retrieval: Foundry calls the Vertex AI RAG Engine through a secure API gateway; Spanner returns the best passages and diagram descriptions.
- Step 7, the governed answer: a foundation model such as Gemini or GPT writes from that context; Foundry checks it before it is shown.
- Beyond troubleshooting: for configuration tasks it pre-generates automation code blocks, checks each block for errors before the engineer sees it and suggests reusable libraries.
Each platform does the job it does best
Tap a job to see who owns it. Google Cloud does the data-heavy work, Microsoft AI Foundry runs the governed agent, and the client's engineering software is where engineers work.
- Set by the brief: the design had to span Google Cloud and Microsoft environments and meet enterprise-grade identity governance and security requirements.
- Google Cloud for retrieval: native ingestion, Gemini parsing for diagrams and Cloud Spanner vector storage that scales to large engineering datasets, on managed services rather than custom vector database infrastructure.
- Microsoft AI Foundry for governance: identity through Entra ID, safety enforcement, agent lifecycle management and orchestration of the retrieval call, in one runtime.
Inside the co-pilot
Product screens with sample data. Tap any screen to see it full size.
Experts back on engineering
Measured after rollout. Drag each slider from before to after.
- Expertise that stays: as experienced engineers retire or move on, what they knew stays searchable.
- Every document, everyone: each file added to the knowledge base is instantly available to every engineer.
- Central control: behaviour and safety rules change in one place, without redeploying.
- Repeatable: the same grounded co-pilot can be built on any organisation's own documentation and platforms.
Expert knowledge, available to every engineer.
A governed co-pilot grounded in your own documentation, diagrams included.
- 01Your business objective
- 02The technical context
- 03The right next step