Current research areas.

EVA-I conducts internal work on techniques that make it possible to run artificial intelligence assistants entirely on the user's hardware, without relying on cloud inference or training infrastructures. The areas below describe the domains currently under study.

Efficient local inference on Apple Silicon

Study of quantization techniques (4-bit, 6-bit, mixed), unified memory allocation on M-series chips, and optimization of inference chains under MLX and Core ML, in order to achieve response times compatible with conversational use on standard user hardware (16 GB RAM).

Personalization through local, reproducible fine-tuning

Adaptation of open language models via LoRA (Low-Rank Adaptation) methods directly on the user's device, without externalizing training data. The aim is for a user to be able to reproduce and control the entirety of their own personalization.

On-device memory for agents

Persistent vector memory architectures stored locally, semantic indexing of user content, relevant contextual recall without transmission to third-party services.

Interface layer between user and cloud models

Protocols enabling a local agent to trigger, explicitly and at the user's request, queries to cloud models of their choosing, isolating the data sent from personal context unnecessary to the query.

European compliance by design

Study of the implications of EU Regulation 2024/1689 (the "AI Act") and of the GDPR on the design of edge AI solutions, and identification of architectural choices that reduce or eliminate personal data processing obligations at the source.

Publications

Open contributions (code, derived models, datasets, papers) will be released as work progresses through open channels (Hugging Face, GitHub) under permissive licences compatible with commercial use. This page will be updated with each publication.

Last updated: 16 May 2026.