Each colour maps a published value to the exact source passage used to support it. Only bounded excerpts are public; administrators can inspect the complete captured source.
Title
DACH insurance group runs an on-prem Qwen3-32B LLM for claims and underwriting staff with no data leaving its own infrastructure
…region had a working AI pilot its own lawyers would never let into production. The answer was an on-prem LLM: Qwen3-32B on two GPUs in the client's own building, with no outbound internet access. Three-year cost fell 81%, the hardware paid for itself in 3.9 months, and blind…
Description
A large Germany-headquartered insurance group supervised by BaFin moved an AI pilot to production by running Qwen3-32B (AWQ 4-bit, vLLM) on two NVIDIA L40S GPUs in its own server room, with no outbound internet access. An internal RAG platform serves a copilot for claims and underwriting staff plus batch summarization and classification, with SSO-bound role-based access and a per-request audit trail. Built by Valletta Software.
…DPIA, DORA assessment, penetration test and legal sign-off done before go-live The client is a large insurance group headquartered in Germany, writing business across the DACH region and supervised by BaFin , Germany's financial supervisory authority. The company is not named, and no p…
…ontrolled deployment process. Retrieval over the documents that actually matter The model is reached through an internal RAG platform, not a raw chat box, indexing underwriting manuals, claims guidelines, policy and product documentation, the knowledge base and se…
Company
Valletta Software (client: unnamed DACH insurance group)
…ent's own infrastructure. How long does it take to move an AI pilot on-premise? This one took six weeks from the first architecture workshop to production, with four Valletta engineers plus the client's QA and compliance specialists part-time. Running the compliance work in parallel with the build, rather than after it, i…
Country
Germany
…DPIA, DORA assessment, penetration test and legal sign-off done before go-live The client is a large insurance group headquartered in Germany, writing business across the DACH region and supervised by BaFin , Germany's fi…
Industry
Insurance
…m LLM Case Study Their Lawyers Banned the Cloud. The On-Prem LLM Cost 81% Less. A large insurance group in the DACH region had a working AI pilot its own lawyers would never let into production. The answer was an on-prem LLM:…
Problem
The group had run a proof of concept against an external frontier API, but employees had to strip personal data out by hand and paste in only fragments. Legal and information security refused production use with real documents because of GDPR obligations, BaFin expectations and DORA requirements, client non-disclosure agreements and an internal ban on sending documents or prompts to external AI providers.
…and underwriting staff saved real time. The problem was how they had to use it: employees stripped personal data out by hand and pasted in only fragments, because nothing else was permitted. That is a fine way to run an experiment and a useless way to run an insurance o…
Solution
Qwen3-32B Instruct, quantized to AWQ 4-bit and served with vLLM behind an OpenAI-compatible internal endpoint, runs on two NVIDIA L40S GPUs in the client's own server room in Germany with no direct outbound internet access. It is reached through an internal RAG platform over underwriting manuals, claims guidelines and policy documentation, with answers citing internal documents, SSO-bound role-based access and a per-request audit record.
…e Solution: Four Layers, All Inside the Building Serving: Qwen3-32B on two GPUs The production model is Qwen3-32B Instruct, quantized to AWQ 4-bit and served with vLLM behind an OpenAI-compatible internal endpoint, tensor parallel across two GPUs. That endpoint matters more than it sounds: internal applications talk to it exa…
Business value
The vendor reports, from actual request and token volume after launch, a three-year cost of 75,580 euro in-house against 392,400 euro for a comparable external API (about 81% lower) and hardware payback in 3.9 months. In a blind review of 600 tasks by the client's experts, 93.4% of on-prem answers were accepted against 94.6% for the frontier model. The system serves around 42,000 inference requests a day.
…a month , including power and cooling, monitoring, backup and hardware support Over three years, 75,580 euro in-house against 392,400 euro on the API : a saving of 316,820 euro, roughly 81% Hardware payback at current volumes: 3.…
…n-house against 392,400 euro on the API : a saving of 316,820 euro, roughly 81% Hardware payback at current volumes: 3.9 months Cost was not why this project happened, and it is worth saying so, because the…
…ulness than the external model. The Results: 81% Cheaper, and Actually Approved The system serves around 42,000 inference requests a day at a peak concurrency of 34. An interactive copilot for claims and underwriting…
AI capabilities
Large Language Models, Retrieval-Augmented Generation
…ontrolled deployment process. Retrieval over the documents that actually matter The model is reached through an internal RAG platform, not a raw chat box, indexing underwriting manuals, claims guidelines, policy a…
Technology
Qwen3-32B, vLLM, NVIDIA L40S
…a commercial API, so the rest of the estate needed no bespoke integration work. The hardware is a dedicated inference node with two NVIDIA L40S 48 GB cards, 512 GB of RAM, mirrored NVMe storage and redundant power, in the client's own…
Deployment model
On-premise
…DIA L40S 48 GB cards, 512 GB of RAM, mirrored NVMe storage and redundant power, in the client's own server room in Germany, with no direct outbound internet access. Weights, containers and updates arrive through a controlled deployment process…
Deployment options
On-premise
…DIA L40S 48 GB cards, 512 GB of RAM, mirrored NVMe storage and redundant power, in the client's own server room in Germany, with no direct outbound internet access. Weights, containers and updates arrive through a controlled deployment process…
Applicable regulations
General Data Protection Regulation (GDPR), DORA
…with real documents, legal and information security said no, for four reasons: GDPR obligations covering policyholder and claimant personal data BaFin expectations and DORA requirements for ICT risk management Non-disclosure…
…four reasons: GDPR obligations covering policyholder and claimant personal data BaFin expectations and DORA requirements for ICT risk management Non-disclosure agreements with corporate clients, forbidding any transfer of th…
Implementation approach
Six weeks from the first architecture workshop to production, with about 980 hours of engineering effort and a team of four plus the client's QA and compliance specialists part-time. Information security and data protection joined the first architecture workshop, and the compliance work ran in parallel with the build.
…, information security and data protection teams. How It Was Built in Six Weeks Six weeks from architecture workshop to production, about 980 hours of engineering effort, and a team of four: an ML and AI engineer, a backend engineer, a DevOps and security engineer and…
…tem first and hand it to compliance at the end. That ordering killed the pilot. Here the first architecture workshop included information security and data protection, and the constraints were fixed before any model was chosen: no data egress, no…
Headline outcome
…VIDIA L40S GPUs, served with vLLM behind an OpenAI-compatible internal endpoint 81% lower three-year cost: 75,580 euro in-house against a projected 392,400 euro on an external API 93.4% accepted answers against 94.6% for the frontier model, and better citatio…
Use case type
Knowledge search
…ontrolled deployment process. Retrieval over the documents that actually matter The model is reached through an internal RAG platform, not a raw chat box, indexing underwriting manuals, claims guidelines, policy and product documentation, the knowledge base and se…
