AI engineering

Customized AI systems
for regulated companies.

AI engineering for SMEs, law firms, medical practices and industry. On-premise LLM (Llama, Mistral) on your own hardware, RAG with data sovereignty, agents integrated directly into your applications. No mere API wrapper, no cloud lock-in.

Free initial consultation → Industry Solutions
GDPR + on-premise possible Integrated in Django From the Regensburg region
On-Premise LLM server in a customer data center
On-Premise · Customer hardware

Engineering notes

How we build — notes from the engineering bench.

Documented architecture decisions, benchmarks from production deployments and engineering patterns we reuse. No generic tutorials, no marketing whitepapers — only the questions we had to answer on real customer projects.

Note 01 · RAG

How we evaluate a RAG system before it goes live

Our internal eval harness: 7 metrics, three ground-truth sets from real customer questions, automated regression on every model swap. What we measure — and why “works in the demo” isn't enough.

pgvector recall@k ground-truth
Lesen → Note 02 · Inference

Llama 3.1 70B on a single L40S — what we measured

Quantization (Q4/Q8), batched inference, vLLM configuration. Where the latency knee sits, when a second GPU is worth it, and which token sizes halve throughput. Real numbers from a production deployment.

vLLM P95 latency quantization
Lesen → Note 03 · Architektur

pgvector next to the Django ORM — pattern and benchmark

Why we put vector search directly in Postgres instead of running a separate Pinecone or Weaviate instance. Indexing strategy, HNSW tuning, recall comparison. With the ORM mixin we reuse in every project.

HNSW Django ORM Postgres
Lesen →
Updated with every new production deployment. Alle Engineering Notes →

Our AI stack

Open-source LLMs, integrated in Django, on your hardware.

We rely on production-ready, open components. Models and tools interchangeable, no vendor lock-in, no hidden cloud dependencies.

Models Llama 3.1 · Mistral · DeepSeek Open-source LLMs, quantisable (Q4/Q8) for a lean hardware footprint. Commercial APIs where sensible — never for personal data.
Inference vLLM · Ollama Production-grade serving on your own GPU (RTX 4090, L40S or larger). Batched inference, dynamic quantization, persistent sessions.
Retrieval (RAG) pgvector · Postgres Vector search directly alongside your business data — no separate vector database, no additional data copy.
Voice & Image Whisper · XTTS · Vision-LLM Speech transcription (DE/IT/EN), natural TTS and vision models for documents, plans, photos — everything possible on-premise.
Orchestration Django · Tool-Calling Agents live in your Django app with direct ORM access. Tool calls via vetted whitelists, audit log included.
Compliance DSGVO · EU AI Act Risk classification, technical-organizational measures, order processing, and transparency documentation — from the very beginning.
AI engineering in detail →

Engagement models

Four ways to work with us.

From AI strategy audit to fully operated on-premise cluster. Each step with a fixed price after the strategy conversation, without hourly trap.

01

AI Strategy & Audit

Workflow analysis, use case assessment, architecture recommendation. Including EU AI Act classification of your existing systems and a written roadmap.

Fixed price · upon request

02

Pilot Implementation

A production-ready AI agent or RAG prototype, embedded in your Django application. Weekly demos, iterative until go-live.

Fixed price · upon request

03

Production rollout

Complete on-premise setup with own hardware, vLLM/Ollama stack, monitoring, GDPR documentation, and maintenance contract from day one.

Fixed price · upon request

04

Managed AI

Ongoing operation of your AI systems: model updates, GPU maintenance, vLLM tuning, compliance reviews and extended availability on request.

Monthly · upon request
Free initial consultation →

FAQ

The most common questions before the call.

We develop AI systems within your Django application and run models on your own hardware. The same team handles development and operation.

That is exactly what we are here for. We analyse your workflow, build, deploy and maintain the system — including model updates and ongoing operation. You need no in-house AI expertise: you get documentation, a clear maintenance path and a direct contact person, no call centre.

After the strategy audit, a pilot delivers a production-ready use case in 8–12 weeks — with weekly demos. No big bang after a year: you see early whether it pays off and decide after every 2-week sprint whether we continue.

Then we tell you — in the strategy audit, before you invest heavily. We honestly assess whether a use case brings real ROI. Better a clear “not worth it here” than an expensive project with no impact. The audit fixed fee is credited against a subsequent pilot.

Let's outline your AI solution.

30 minutes on the phone. We analyze your workflow and honestly tell you whether AI brings ROI for you — and if so, with which architecture (agent, RAG, on-premise or cloud). No AI hype, no sales pitch.

Free initial consultation

+49 941 20 90 28 62 [email protected]

Reply within 24 hours on business days · Mon–Fri 9–18 · CODLAB · St.-Jakob-Str. 6, 93161 Sinzing
Written fixed-price quote GDPR · servers in the EU, in Germany on request Direct developer · no call center