What we built, and what it changed.
Pexon publishes AI engineering case studies from delivery inside industrial companies, banks and asset managers. Each one names the starting position, the system built, the trade-off taken and the measured change. Customers are described by category, size and regulatory context rather than named, because that is what the approvals cover.
Described, not named. Measured, not estimated.
Every engagement below ran in production inside a real company. Each case study states the starting position, the system that was built, the trade-off that was taken and the change that was measured, and links the offer it proves so you can see what the same work would cost you.
The customers are described rather than named. The approvals we hold were given to Pexon Consulting GmbH for its own site, and a customer who agreed to appear there has not thereby agreed to appear here, so we describe the category, the size and the regulatory context instead. That is not a hedge. For a buyer deciding whether a reference is relevant, a German cooperative IT provider serving more than 100 banks under BaFin supervision carries more information than a logo does.
Where a figure appears, it came out of the engagement. Where no figure exists we say what changed in words rather than inventing a percentage, which is why some of these read less impressively than the case studies you are used to.
Case studies
- Model Versioning and FinOps at a Regulated Asset ManagerThree parallel model tracks with half-yearly upgrade windows, a Terraform module library, and Kubernetes right-sizing that keeps bank APIs off GPU nodes.
- Self-Hosted RAG Platform for 800+ Cooperative BanksSemantic chunking, pgvector with HNSW indexing and OpenShift elasticity, built for strict tenant isolation under BaFin and KRITIS obligations.
- End-of-Line Visual Inspection at an Engine ManufacturerCustom CNNs on Databricks with dual detection for known defects and unknown anomalies, feeding a self-service WebGUI with real-time alerting on the line.
- Permission-Aware Maintenance Assistant at an Engine ManufacturerPermissions computed per data node, multilingual embeddings, a deliberate smaller-model choice. Response time down 30%, first-fix up 40%, 5,000+ users.
- Multi-Agent Standards Assistant at an Automotive SupplierA LangGraph multi-agent system answering DIN and ISO questions for 300+ engineers, with pseudonymised tracing and a published accuracy split by query type.
- Product Support Assistant at a Transformer ManufacturerHow a transformer manufacturer cut research per technical enquiry by over 90% with an LLM-independent assistant in its own Azure tenant, at EUR 3-5 per user.
Common questions
Why are your case studies anonymous?
Because the approvals we hold cover an anonymous description, not a logo. A case study naming a customer without written consent for this specific site is a legal problem, not a marketing win. Each one carries the industry, the company size and the regulatory context, which is what a comparable buyer actually matches against.
Are the numbers in these case studies measured or estimated?
Measured, and taken from the delivery record. Where no hard figure exists we say what changed qualitatively instead of inventing a percentage. If a case study shows a number, that number came out of the engagement rather than a marketing workshop.
Can I speak to one of these customers as a reference?
Sometimes, and it depends on the customer. Named reference calls are arranged case by case and need the customer's agreement first. Ask on an architecture fit call and we will tell you honestly which engagements have a customer willing to talk.
Not a sales call. An architecture call.
Thirty minutes with the architect who would actually run the engagement.