Local inference.
Open models.
Reproducibility.

Platform: Apple Silicon.

EVA-I's solutions target platforms equipped with Apple Silicon chips (M1 and subsequent generations), which provide the hardware capabilities required to run large language and vision models locally, with response times compatible with conversational use.

Choosing a single platform makes it possible to guarantee execution quality without multiplying the technical support surfaces the company could not maintain to the same standard.

Inference frameworks.

Local execution relies on publicly available frameworks: Apple MLX and MLX Swift for language models, Whisper for transcription, Core ML for Apple Neural Engine hardware acceleration, and vector inference libraries for semantic search.

No proprietary framework or opaque dependency is introduced into the inference chain. All components can be audited, updated and replaced.

Models: open ecosystem.

The models used come from the open ecosystem (Mistral, Llama, Qwen, Whisper and their derivatives) under permissive licences compatible with commercial use. No proprietary model is embedded.

Personalization: local, reproducible fine-tuning.

The company maintains a local fine-tuning pipeline, validated on user hardware (Apple Silicon M-series, 16 GB). Fine-tuning is performed via LoRA (Low-Rank Adaptation) on documented, reproducible configurations, with no need for cloud training infrastructure.

The goal is for a user or partner to be able to reproduce the personalization on their own machine, with their own data, without dependency on a remote service.

User data.

No user data is transmitted to EVA-I in the nominal operation of the distributed solutions. The company has no collection channel. Personalization is carried by the device and is never re-aggregated for collective training purposes.

Distribution.

Solutions are distributed through the official channels compatible with the conditions of the platform publisher. Binary signing and notarization guarantee the publisher's identity and the integrity of the code delivered.