Life Autopilot

What it does

Life Autopilot is a persistent Collaborative Partner for students and busy people who need help coordinating commitments across time, place, preparation, and communication. It works beyond a standard chat loop: it observes structured context, retrieves relevant memory, makes a bounded decision, and can take a lightweight action.

The problem

A calendar knows that an event starts at 14:00, but it does not know when a person should prepare, leave, re-plan, or ask for help. Generic reminders also do not learn from a user's preferences and past feedback. This creates missed classes, late errands, and abandoned personal goals.

The solution

Life Autopilot combines schedule events, travel context, curated campus knowledge, and user-scoped memory. It can recommend preparation, calculate a late-arrival consequence, send a Gmail notification, save a selected suggestion to Google Calendar, or escalate when a destination or route cannot be grounded.

A nutrition-student scenario demonstrates the companion learning preferences for outdoor cycling, resolving timetable location N204 through a curated campus graph, generating a structured multi-option fitness suggestion with Gemini, answering a follow-up from stored suggestion context, and saving the selected option to Calendar. A separate supplier-errand scenario demonstrates preparation, movement-aware re-planning, quantified delay, and notification.

Why it matters

The companion reduces coordination friction by remembering how the user works and turning context into useful next steps. Personalization is bounded and explainable: profile facts include a user ID, timestamp, source, and feedback or confidence state. The campus graph uses curated facts and does not infer identity or occupancy from device density. Unknown or low-confidence destinations are escalated rather than invented.

AI and Google Cloud

Gemini 3.5 Flash is accessed through Vertex AI using the Google GenAI SDK. Gemini synthesizes structured companion suggestions and contextual notification copy. Deterministic application code validates model output and controls timing arithmetic, schema safety, route grounding, idempotency, user isolation, and evolution safety gates.

Cloud Run hosts the FastAPI backend. Firestore stores user-scoped profiles, timetable context, suggestions, commitments, audit events, and offline evolution records. Google Calendar and Gmail are integration actions when OAuth is configured. The project also provides deterministic local fallbacks for reproducible testing.

Bounded self-evolution

The agent records structured outcomes rather than private chain-of-thought. An offline workflow evaluates representative scenarios, identifies failures, proposes a candidate policy, scores it against the active policy, and requires explicit promotion. Runtime requests never silently rewrite production behavior.

How Codex was used

Codex was used to shape the project around the Collaborative Partner track, implement and test the FastAPI services and integrations, diagnose malformed Gemini output, build personalization and self-evolution workflows, create the mobile shell, and prepare reproducibility and demo documentation.

Reproducibility

Clone the repository and follow README.md for local fallback mode, Vertex/Firestore mode, OAuth integrations, Cloud Run deployment, tests, and scenario scripts. The hosted Swagger UI is available at the linked demo URL.

Built With

• cloud-run
• fastapi
• firestore
• flutter
• gemini-3.5-flash
• gmail-api
• google-calendar-api
• google-genai-sdk
• vertex-ai