ALI: Project Brief

Summary

ALI is a locally hosted AI agent that automates tasks on a desktop computer. It runs entirely on in-house hardware, so no data is sent to a cloud AI provider. It is in active development, with the current focus on getting a keyboard-only Outlook email workflow to run reliably.

Objective

Automate repetitive desktop tasks using a fully local AI, with no cloud dependency and no data leaving the machine.

Approach

• Runtime: llama.cpp, on a Vulkan backend (via a TurboQuant fork).
• Hardware: a single AMD RX 9060 XT with 16GB of VRAM. This is the main constraint on model size.
• Models: a two-model design. A Gemma 12B model carries out actions, and a small E2B model handles planning.

Current status

In development. The test workflow is composing and sending an email in Outlook using only the keyboard: To, Subject and Body navigation, then send.

• Completed: a semantic loop detector, which stops the agent when it is repeating itself without making progress, with its false positives fixed over several iterations.
• Completed: a dedicated driver for the message body, which behaves differently from the other fields.
• In progress: getting the full compose-and-send sequence to run reliably from start to finish.

Risks and constraints

• Reliability. The agent can get stuck repeating actions. Mitigated by the loop detector, which is still being refined.
• Hardware ceiling. 16GB of VRAM limits how large a model can run locally.
• Scope. So far only a single workflow has been tested.