DIRSCHL Academy · Infrastructure Deep Dive
On-Premise AI, Local LLMs & Hybrid Architectures
What can local language models truly achieve on your own hardware? We compare open-weight models, local media tools, and cloud APIs for your specific tasks – focusing on performance, data control, and total cost.
Pre-system review, one-day infrastructure workshop, and concrete architecture recommendation. On-site or remote.
Real-World Benchmarks
Measuring quality, latency, throughput, memory requirements, and parallel operation for specific tasks.
No API Token Costs
Local transcription, media processing, and inference – with transparent hardware, energy, and operating costs.
Pre-Checked
Hardware, network boundaries, security requirements, and typical workloads are checked in advance.
Reality Check for Your Infrastructure
Local, Cloud, or Hybrid – Decided Based on Your Requirements
Companies want to control sensitive information and limit ongoing API costs. At the same time, models that are too small, unsuitable hardware, or missing operational processes can quickly lead to poor quality and low acceptance.
Before the workshop, we review your hardware, IT security requirements, data classifications, and prioritized workflows. During the workshop, we compare suitable local models and tools with cloud services using a representative test set.
Local processing can reduce data outflows, but it is not automatically secure or data protection compliant. Access protection, logging, updates, legal basis, and operational responsibility remain part of the overall architecture.
Technology-Agnostic, Not Local LLM Dogma
Ollama, LM Studio, MLX, Linux GPU servers, or secured cloud APIs: The decisive factors are the task, data, quality objective, team, and operations.
Metrics Over Gut Feeling
A joint benchmark reveals where local systems suffice, where optimization helps, and when a cloud solution remains more economical.
Curriculum
The Training Modules in Detail
The weighting is adjusted after the pre-review based on hardware, security level, number of users, and workloads.
Module 1
Local Toolchains Without API Token Costs
- Media Workflows: Utilizing Whisper, ffmpeg, and local processing for transcription, conversion, extraction, and structured follow-up processes.
- Local Inference: Comparing Ollama, LM Studio, and MLX as desktop, server, or API building blocks.
- Hardware Matching: Relating RAM, VRAM, Unified Memory, Quantization, Context Size, and Parallelism to real user numbers.
No API token costs does not mean free: Acquisition, energy, administration, and updates are factored in.
Module 2
Reproducible Comparison: Local vs. Cloud
- Open-Weight Models in Test: Evaluating suitable Llama, Qwen, Mistral, or comparable models using a common set of tasks.
- Performance Profile: Measuring response quality, source relevance, latency, tokens per second, memory requirements, and parallel operation.
- Task Mapping: Separating confidential, frequent, or clearly defined tasks from complex cloud and multimodal requirements.
“Hallucination-free” is not a serious promise. Instead, we define test cases, source verification, and release steps.
Module 3
Security, Data Protection & Hybrid Architecture
- Controlled Data Paths: Design network boundaries, authentication, roles, logging, and update processes for local services.
- Hybrid Routing: Classify data by protection requirements, preprocess locally, and integrate external APIs only as needed.
- TCO Comparison: Compare owned hardware, rented servers, administration, utilization, and API consumption over a realistic period.
Important Note: The workshop provides technical and organizational decision-making foundations but does not replace individual legal or data protection advice. Binding assessments are made in consultation with your legal counsel and data protection officers.
What You Get After the Workshop
Benchmark Report
Measurements and quality assessment of the tested models on your representative tasks.
Architecture Sketch
Recommended components, data paths, network boundaries, and interfaces for local or hybrid setups.
TCO Model
Comparison of hardware, operations, administration, utilization, and cloud consumption.
Action Plan
Prioritized steps for proof of concept, security review, pilot operation, and scaling.
Key Details & Investment
Process & Preparation
- Analysis of hardware, data classes, and security requirements
- One-day workshop with live benchmark
- System recommendation, architecture sketch, and action plan
Target Audience
IT management, CTOs, data protection officers, software architects, and executive management in data- or security-sensitive areas.
Conducted on-site or remotely.
Investment
From €2,800 plus VAT.
Team flat rate including preliminary system review, custom agenda, one-day workshop, and follow-up.
Your Trainer & System Architect
Johann Dirschl
Over 30 years of experience as a lecturer and software developer. His expertise includes APIs, hardware control, local media processing, server workflows, and modern language models. He develops and operates his own applications, plugins, and AI-powered systems.
“We don't sell a dream of a small local model. We test your tasks on suitable hardware and develop an architecture that honestly weighs data control, quality, and total costs against each other.”
Let Us Conduct a Preliminary Infrastructure Review
Tell us about your hardware, user count, data classes, and two to three typical workloads. You will receive an honest assessment of whether a local, cloud, or hybrid architecture is the most sensible next step.