Secure On-Prem AI / Automation Platform
Design and implementation of an \"AI behind the firewall\" fabric combining document classification, extraction, routing and decision workflows on heterogeneous hardware for regulated clients.
The brief was document automation for organisations that cannot send a document, a prompt or a model output outside their own network. The answer was a platform that runs inference where the documents already are.
Documents enter through a hardened intake: scanned paper, PDFs and images are OCR’d with GPU acceleration, native PDF text is extracted directly, layout and tables are recognised, and the result is a classified, searchable document with embeddings for semantic search. From there, extraction and routing turn what used to be a manual inbox flow into structured events that downstream decision workflows and people act on, with human approval wherever the effect leaves the system.
Underneath sits the same sovereign platform the studio runs for itself: Kubernetes clusters by GitOps on mixed hardware, secrets in Vault, private model serving, object storage and backups. It operates in the client’s data centre or hosted in Berlin, with the same code either way.