Runs local. Stays local.
On-premise AI for medical practices, law firms and tax advisors. It handles correspondence, case files and follow-ups on its own — and client and patient data never leaves your systems. § 203 StGB compliant by design.
The situation
The desk keeps getting fuller. Correspondence, follow-ups, receipts, case preparation — client time gets eaten up by all of it. The work hasn’t become harder, just more of it. And exactly where you should be there for your clients or patients, you’re sitting at the desk sorting paper.
The open position stays open. Assistants, front desk, medical assistants, tax clerks — posted for months, hardly any qualified applicants. The work nobody takes ends up with the owner. The bottleneck is demographic and therefore structural — it won’t disappear.
And ChatGPT isn’t allowed to read along. § 203 StGB effectively forbids it. Putting client or patient data into a US-based AI cloud isn’t a privacy question — it’s a criminal offense, regardless of how GDPR-compliant a vendor advertises itself. Even German cloud AI doesn’t solve this legally as long as you don’t have documented consent for every single client on file. In practice, that’s where every conventional AI tool fails.
The answer
The assistant handles tasks end-to-end: drafting emails, summarizing case files, preparing appointments, pre-checking invoices. It only asks you where you have to decide.
01Everything runs on hardware in your office. No data outflow, no cloud dependency, no per-seat subscription. Once it's set up you can unplug the internet — the assistant keeps working.
02We tune the assistant to your case files, your templates and your phrasing. It responds in your firm's tone, not in the tone of a Silicon Valley chatbot.
03No cloud. No US servers. No data outflow. Runs local. Stays local.
Services
From strategic analysis to ongoing operation of your AI assistant — all from one partner, all aligned with your professional code of conduct.
A local AI assistant for your firm or practice — from hardware selection and the right language model to rollout and staff training.
Your assistant summarizes briefs, finds passages across your entire file base and drafts letters in your tone — all locally, without cloud forwarding.
Process analysis, opportunity identification and a legal feasibility check — before you invest.
Websites that match your professional code of conduct — GDPR-clean, no tracking traps, no US dependencies.
How it works
The idea in three layers, all under your control — together they form an AI assistant that never leaves your premises.
A compact local device — Apple Silicon or NVIDIA GPU, sized to your firm or practice. It sits in your server room or office. We handle setup, configuration and security hardening.
We install open language models (Llama, Qwen, Mistral) on your hardware. No API calls, no external inference. The models have documented weights and can be audited independently — and they power the assistant your team uses day to day.
Your documents are indexed and processed locally. No upload to third parties. No model training on your data. No telemetry. Pull the plug from the internet — the assistant still works.
Which hardware you need depends on your firm or practice size, your use cases and your expected response behavior. We make that decision together during the introductory call.
Investment
We work project-based instead of a monthly retainer. You pay for a clearly defined outcome — not for idle capacity.
Two-week structured process analysis with legal classification and a concrete implementation recommendation.
A full local AI assistant for one core application — hardware, model and integration for a single firm department.
Firm-wide AI assistant for multiple departments and use cases — with high availability and backup.
Ongoing operation (optional): €850 – €3,500 per month across three tiers (Basic / Business / Premium) — response times 8 h / 4 h / 1 h, model updates every 3–6 months, quarterly firm-advisory session. Additions outside the packages are billed transparently: consulting €180/h, engineering €140/h.
All prices exclude statutory VAT.
About the founder
My name is Dr.-Ing. Sven Rost. I hold a degree in mechatronics and a doctorate in robotics from two of Germany’s leading technical universities — with a foundation in control engineering, system identification and automation.
Over the past 15+ years I’ve held international leadership roles in four countries — as institute director, vice-rector and strategy advisor. I’ve built cooperation networks, raised seven-figure third-party funding and advised at government level. I currently also lead strategic partnerships and business development at an international strategy consultancy.
I founded the company because the very professions that would benefit most from AI relief — tax advisors, attorneys, physicians — are the ones legally least able to use the technology. The answer isn’t another cloud service with marketing promises, but sovereign infrastructure that runs on your hardware, under your control, by your rules.
Philosophy
AI should relieve your staff — not replace them. Every assistant is built together with your team and tailored to your concrete workflows.
Questions & answers
Our focus is on professionals bound by § 203 StGB confidentiality — specifically tax advisors (§ 62a StBerG), attorneys (§ 43e BRAO) and physicians. There, using mainstream cloud AI ranges from legally problematic to outright prohibited.
Other sectors with strict confidentiality requirements also benefit from on-premise AI: auditors, notaries, patent attorneys, psychotherapists, mid-sized research departments and companies with high IP-protection needs. Reach out — in the introductory call we’ll assess whether our approach fits your situation.
It depends on three factors: the number of concurrent users, the type of documents (PDF briefs vs. scanned files), and the response speed you expect.
Rough orientation:
We finalize the exact configuration after the use-case analysis. We don’t resell hardware — you purchase directly from the manufacturer or reseller of your choice.
A typical project runs eight to twelve weeks from contract signing:
For the AI audit as a standalone upfront project, expect two weeks of turnaround from kickoff.
Yes — and that’s not a marketing promise, it’s a property of the system design.
The assistant runs on hardware in your rooms. There are no API calls to external providers, no training-data transmission, no telemetry. You can physically disconnect the device from the internet and the assistant still works.
On top of that come standard on-premise practices: encrypted storage, SSH hardening, detailed audit logs of every query, role-based access. On request we draft a Data Protection Impact Assessment (DPIA per Art. 35 GDPR) and a data processing agreement with a § 203 addendum.
The difference to cloud AI is structural: with a US vendor you have to collect consents and trust the vendor. With an on-premise solution your data never leaves your infrastructure in the first place — the consent question simply doesn’t arise.
The introductory call runs 45 minutes, is free of charge and non-binding. We get to know how you work and identify together where AI would make a measurable difference.
Within five business days you receive a written assessment from us with:
If you want to proceed after that assessment, we sign an offer and start. If not, you still have a solid basis — feel free to use it for conversations with other providers.
We work with a two-stage safeguard model instead of a blanket money-back guarantee — because the latter wouldn’t be honest for hardware-based projects.
First: proof-of-concept before the full investment. If you’re unsure, start with the AI audit (€3,900) or a two-week proof-of-concept on loaned hardware. Only then do you decide on the full project.
Second: fixed price instead of hourly billing. Our project prices are fixed. If we’ve underestimated the effort, we carry the risk — not you.
Hardware you’ve already purchased stays yours and keeps its value even outside our setup. The models are open source, the code is documented, the documentation sits with you. You’re bound to us neither technically nor contractually.
We consistently rely on open-source components and documented interfaces — no proprietary SaaS dependencies in the production path.
Models & inference. Large language models from the open ecosystem (including Llama, Mistral, Qwen), run locally via runtimes like Ollama or vLLM. Model choice and quantization depend on your hardware — Apple Silicon (M-series) for compact firm setups, NVIDIA GPUs for larger context windows and parallel assistant roles.
Operation & access. Cloudflare Tunnel for external access, so no open ports are exposed on the firm’s network. Authentication via TOTP-based 2FA, passwords hashed with bcrypt, role-based permissions, complete audit logs of every query. Daily encrypted backup, regularly tested restore.
Portability. The entire stack runs containerized (Docker) and on standard Linux hardware. Code, configuration and documentation sit with you — neither models nor infrastructure are tied to a proprietary vendor. If you end the engagement with us, the system remains fully functional and can be taken over by any service provider with Linux expertise.
Concrete versions, quantization levels and hardware sizing are determined in the audit phase against your real workload — not upfront from a catalog.
Introductory call
We get to know how you work, identify the biggest levers together, and within five business days deliver a written assessment — with a concrete recommendation and transparent pricing.
Pick any free 45-minute slot. We meet by phone or video — your choice, we call you.
For a fast follow-up or scheduling question — no app install needed, both links open the respective chat directly.
Tell us about your situation in two or three sentences — we'll get back to you personally within one business day. Or write directly to contact@merlon.ai.
