DE
Menu
DE

Pahl Digital Art
Kilchberg ZH · Switzerland

Local AI, LoRA training & LLM integration

AI needs to fit both the task and the device. Pahl Digital Art works on integrating local language models into applications, drawing on Swiftli Smart HR, adapter training and optimisation on Apple Silicon.

Start a project

Start with a specific function

We establish what the model receives, what result is needed and how the application checks that result. This includes cases where a reliable answer is not possible. A language model supports a workflow; understandable rules and human oversight remain important.

First establish whether the task works locally

A useful starting point is a focused test with representative inputs on the intended device. We check output format, result quality, failure cases, response time and memory use. This helps establish whether a base model is sufficient, an adapter would help or another method would be more suitable.

A model on one Mac and a shared local server are different operating models. Our practical experience includes integration on Apple Silicon. Server operation also requires consideration of the network, access rights, concurrent use and ongoing support.

Consider adapters, quantisation and caching together

LoRA training uses additional adapters to adapt a model to selected tasks. Quantisation can reduce memory requirements. Pre- and post-caching can avoid repeated work. These techniques affect quality, memory and response time differently, so they need to be evaluated together.

  • Define the task, data format and quality criteria
  • Prepare datasets for adapter training
  • Evaluate LoRA adapters and quantised models
  • Integrate caching into the application workflow
  • Test performance and output quality on the target device

Measure on the hardware people will use

Local LLM integration for Swiftli Smart HR was also explored with a MacBook Air M1 in mind. Output tokens per second are only part of the picture: time to first response, memory use and result quality matter too.

Measurements depend on the model, quantisation, input length, cache state and hardware. A reliable commitment for your project therefore comes from a relevant test on the intended devices.

Common questions

Does local AI keep all data on the device?

Model computation can run locally. Whether all data stays on the device also depends on synchronisation, logging and connected services. The data paths of the entire application need to be considered.

Does every task require adapter training?

No. We first assess what a suitable base model can achieve with clear inputs and output rules. Training is useful when it provides a measurable advantage for the specific task.