Overview
An agentic assistant that automates phone and desktop workflows — scheduling, messaging and data tasks — through LLM tool-calling.
The problem
Repetitive device tasks — scheduling, messaging, moving data between apps — are time-consuming and easy to get wrong. The goal was an assistant that could actually operate the device, safely, not just chat about it.
Approach
- Designed a small set of strictly-typed tools (
send_message,create_event,read_contact) with clear schemas and descriptions. - Implemented a bounded agent loop with step limits, timeouts and per-task cost caps.
- Classified tools by risk and required explicit confirmation for anything irreversible or external.
- Used a small local model for routing and a larger model only for complex steps — keeping runs fast and private.
- Logged every reasoning step, tool call and result for debugging and evaluation.
- Mitigated prompt-injection by refusing to let untrusted content trigger tool calls directly.
Result
A predictable on-device agent that automates real workflows while keeping the user in control of anything consequential — reliable enough to trust with daily tasks.