Explore projects

MobileTransformers is a lightweight, modular framework based on ONNX Runtime for running and adapting large language models (LLMs) directly on mobile and edge devices. It supports on-device fine-tuning (PEFT), efficient inference, quantization, weight merging, and direct inference from merged models. It includes advanced generation techniques like Retrieval-Augmented Generation (RAG) with vector databases and KV-cache with embedding reuse. The framework also provides export scripts for converting custom Huggingface SLM/LLM for on-device deployment with custom PEFT methods.

Comparison of Slimmable Neural Network and Any-precision (quantized neural network) on a use case of human activity recognition on mobile devices. A full implementation of an SNN running on an IoT devices is provided as well.

Wireless ranging with participants engaged in different cognitive tasks.

Study of mobile UI adaptation for news viewing.

Reactions to personality-tailored mobile notifications.

IoT-sensed data with users engaged in different tasks.

Mobile video watching behaviour while performing different activities.

Cognitive load-related physiological signals while playing a game.

Mobile notifications and context dataset.