AI or Classic Algorithm?

The most important question comes first

Not every problem needs AI. A classic algorithm is deterministic, explainable, cheap to run and delivers the exactly correct result when the rules are clear — calculations, optimisations and well-defined processes belong to it. AI plays to its strengths when rules are unknown or too complex: recognising patterns in images, speech or sensor data, making predictions from historical data, or understanding natural language.

Our rule of thumb:
  • Rules known and expressible → classic algorithm (faster, traceable, maintainable)
  • Patterns in large data sets, perception tasks (seeing, hearing, language) → machine learning
  • Both combined → hybrid: AI detects, classic logic decides

We have been building both worlds since 1999. That is why our advice is technology-neutral — your problem determines our proposal, not the buzzword.

Machine Learning, Deep Learning or LLM?

The right AI technology for your use case

  • Large language models (LLM) understand and generate language. Ideal for document analysis, intelligent assistants, knowledge bases (RAG) and automating text workflows.
  • Deep learning (DL) uses neural networks for perception tasks: image and object recognition, camera-based quality control, speech and audio processing. Needs more data and GPU power, but delivers results that classic ML cannot reach.
  • Machine learning (ML) learns patterns from structured data. Ideal for predictions (maintenance, sales), anomaly detection in sensor data and classification. Needs comparatively little data and computing power — and runs on small hardware if required.

Which level is right depends on your data, budget and explainability requirements. We assess this in a free initial consultation — quite often the lean solution is the better one.
LLM integration into your software: AI reads technical documentation, schematics and manuals

LLM Integration Into Your Software

Using language models sensibly — cloud or local

We bring language models into your existing software: chat with your own documents (RAG), automated report generation, semantic search or voice control of applications and exhibits.

Together we choose the right approach:

  • Cloud connection to the large models from OpenAI, Anthropic or Google — best model quality, fast implementation, pay-per-use. Your data passes through the provider's API.
  • Local open-source models (e.g. Qwen, Llama or Mistral, run via Ollama) on your own hardware — full data sovereignty, no ongoing API costs, at the price of somewhat lower model quality and your own GPU requirements.
  • Hybrid: pre-process sensitive data locally, send only non-critical requests to the cloud.

We use both worlds ourselves every day, so we know from practice what the models can really do — and where their limits are.
Computer vision & pattern recognition in production use

Computer Vision & Pattern Recognition

In productive use by us for years

Image recognition at at² is not a promise for the future — it has been everyday business for years. With technologies such as OpenCV and Fraunhofer SHORE we have delivered, among other things:

  • Real-time face analysis (age, emotion, gender) for interactive trade-fair exhibits
  • Face mapping and live filters: national flags, masks and Snapchat-style effects projected directly onto visitors' faces
  • "Krake lebt!" at Heide Park Resort: visitor photos transformed into sea monsters via face recognition — robust in theme-park continuous operation
  • Gesture and touch recognition with TUIO, IR rear-projection sensors, depth cameras, floor lasers and Leap Motion
  • Autonomous navigation: obstacle detection and 3D pathfinding for drones (SPRIND challenge "Fully Autonomous Flight", together with Vectorbirds Airborne Systems)

For industry this means: automated quality control, counting and measuring by camera, real-time process monitoring — with proven technology instead of experiments.

AI on Your Own Hardware

From industrial PCs to microcontrollers

Not every AI needs the cloud — but not every AI fits on every chip. We give you a realistic assessment of which intelligence runs on which hardware:

  • Workstations & industrial PCs with GPU: local LLMs, complex image recognition, multi-camera systems
  • Embedded boards such as NVIDIA Jetson or Raspberry Pi: object recognition and image classification directly on the line — no internet connection, millisecond latency
  • Microcontrollers (STM32, ESP32): lean, specialised ML models (TinyML) for anomaly detection in sensor data, vibration or noise analysis — no language models here, but surprisingly capable pattern recognition

The bridge between our embedded development and modern AI: a model is worthless if it does not run reliably on your hardware.

AI Exhibits

Experiences that were impossible five years ago

AI enables exhibits that simply did not exist five years ago — and we combine them with more than 20 years of exhibit experience in continuous operation. What is possible today:

  • Generative imagery: visitors are transformed into paintings, historical scenes or brand worlds
  • Talking avatars: digital characters that answer visitor questions in real time — on request with a local LLM, entirely without the cloud
  • Generative art: installations that create live visuals from movement, sound or visitor data
  • Photo experiences: the evolution of our greenscreen and face-mapping exhibits with generative AI
  • Voice control: exhibits that listen and respond — multilingual and low-barrier

We have implemented face recognition, face mapping and visitor transformations many times (Heide Park with OpenCV, trade-fair exhibits with Fraunhofer SHORE) — generative AI is the next logical step for us, not a first attempt. GDPR-compliant and robust for trade-fair and theme-park operation, of course.

Data Privacy & EU AI Act

AI you can rely on

AI projects rarely fail because of the technology — more often because of data privacy, compliance and missing maintainability. That is why the following is part of every project from day one:

  • GDPR-compliant architecture: data minimisation, local processing where possible, no storage of biometric data in our recognition exhibits
  • EU AI Act: classification of your project into the risk classes, transparency and documentation duties considered from the start
  • On-premise options: sensitive data never leaves your premises — local models on your hardware
  • Traceability: documented models, reproducible results, clear limits — we do not promise magic

Made in Germany applies to our AI as well: developed, hosted and maintained in Germany.

Frequently Asked Questions About AI Development

Everything about AI projects at at²

What does the development of an AI solution cost? +
Feasibility prototypes start below €3,000 with us. They clarify first whether your use case works reliably with AI before you invest more. Production solutions depend on data, hardware and integration depth — after a free initial consultation you receive an honest estimate.
When is AI worthwhile — and when is a classic algorithm better? +
A classic algorithm is the better choice whenever the rules of your problem can be clearly formulated: it is faster, cheaper and traceable. AI pays off when patterns need to be recognised that cannot be captured in rules — in images, speech, sensor data or historical data. Exactly this assessment is the first step of every consultation with us.
Do we need large amounts of data for machine learning? +
Not necessarily. For many tasks, pre-trained models adapted to your case with a few hundred examples are sufficient. We honestly assess whether your data is adequate in the initial consultation — and tell you if it is not.
Can LLMs be operated privacy-compliant and without the cloud? +
Yes. Open-source models (e.g. Qwen, Llama or Mistral, run via Ollama) run entirely on your own hardware — your data never leaves your premises. Model quality is below the large cloud models from OpenAI or Anthropic, but is fully sufficient for many business applications such as document search or internal assistants.
How long does an AI project take? +
A feasibility prototype typically takes 2 to 6 weeks with us. Production solutions including integration, testing and training take a few months depending on scope. We work iteratively — you see intermediate results from the start.
Do you integrate AI into existing software as well? +
Yes, that is actually the standard case: AI features are integrated into existing systems, machine controls, apps or exhibits instead of rebuilding everything. We have been developing custom software since 1999 — we know the system landscapes AI has to fit into.
How do you ensure the security of AI applications? +
AI software goes through the same security process as any other application at at²: static code analysis (SAST) during development, dynamic security testing (DAST) against running systems, and regular dependency audits. This is especially important for AI projects with many dependencies such as Python packages and ML frameworks.
What is an SBOM and why does it matter for AI software? +
An SBOM (Software Bill of Materials) is a machine-readable list of all components and dependencies in a piece of software. For AI projects with extensive library dependencies like PyTorch, TensorFlow, or Hugging Face, an SBOM is particularly valuable: it makes transparent what is inside your software and enables fast response when new vulnerabilities are discovered.
Is our AI software ready for the Cyber Resilience Act? +
The EU Cyber Resilience Act requires end-to-end cybersecurity for products with digital elements from December 2027 — this also applies to AI components in your products. Our workflow with SAST, DAST, dependency audits, and SBOM generation prepares your AI solution for these requirements.

The Free AI Check

Know where you stand in 30 minutes

Describe your use case — in a free 30-minute call we tell you whether and which AI makes sense for it, what a feasibility prototype would cost, and whether there is a simpler way. Honest, non-binding, without buzzwords.

Request an AI Check Call Us Directly
Heide Park - The Kraken Lives!

Heide Park Resort

Facial recognition in the horror labyrinth

In the horror labyrinth leading to the Krake roller coaster, visitors were photographed on entry - and later reappeared in the scenery transformed into sea monsters, thanks to facial recognition (OpenCV). Robust software running continuously in daily theme park operation.

  • Exhibit
  • Facial Recognition
  • OpenCV
  • Theme Park
  • 24/7 Operation
MediaScreen - TUIO Driver

MediaScreen

Multitouch recognition for MediaSphere

For our long-standing partner MediaScreen, we developed the TUIO touch recognition for the IR rear-projection systems Monkeybook and MediaSphere completely from scratch - including fine calibration and our own calibration software. Since then, the hardware has been rented and sold worldwide with our software.

  • Multitouch
  • TUIO
  • IR Sensing
  • Calibration
  • Partnership
Software development since 1999 - Made in Germany

Made in Germany

Software development since 1999

Developed 100 % in Germany — at our Nuremberg and Kempten (Allgäu) locations. Over 1,170 completed projects, with 82 % of our revenue coming from repeat customers. If you're nearby: stop by for a coffee.