LifeOps

Viewed HACKATHON_DEMO.md:1-69 Viewed ARCHITECTURE.md:1-100

Here is the structured breakdown for LifeOps (README.md):

Inspiration

Modern AI assistants are great at answering questions, but terrible at taking real responsibility. When facing a complex life event—like relocating to a new apartment—users are still forced to manage endless checklists: analyzing lease clauses, timing notice windows, scheduling technicians, reserving movers, transferring utilities, and handling cancellations.

We were inspired to shift the AI paradigm from "Chatbot Advice" to "Outcome Delegation." What if you could simply tell an AI: "Sort out my apartment move and broadband," and have an autonomous agent reliably manage the logistics, respect risk guardrails, verify actual progress, and adapt when real-world plans fall through?

What it does

LifeOps is an autonomous personal operations manager that turns high-level life goals into completed real-world outcomes.

• Outcome Delegation: Takes ambiguous goals (e.g., apartment relocation) and independently scans user vaults (leases, utility bills, notices) to extract structured facts and timelines.
• Dynamic Task Graphing: Automatically breaks down goals into multi-step task dependency graphs (e.g., landlord notices, broadband transfers, mover bookings).
• Deterministic Risk Engine: Executes safe background tasks (e.g., checking slots, drafting emails) automatically, while deterministically pausing for Human-in-the-Loop approval on high-risk financial or legal actions (e.g., paying move deposits).
• API Verification: Queries third-party provider systems to confirm actual completion rather than assuming tool call success.
• Event-Driven Replanning: Detects emergency vendor cancellations (e.g., moving company breakdown), instantly wakes up, searches for alternatives, and replans the workflow autonomously.

How we built it

We engineered LifeOps using a privacy-first, zero-cloud local architecture:

• Agent Hierarchy: Built on Google ADK (Agent Development Kit v2.3.0) with dedicated specialist agents:
  • Orchestrator Agent: Manages global state, phase transitions, and paused workflow recovery.
  • Discovery Agent & Document Agent: Parse lease terms and utility bills to extract structured facts with confidence ratings.
  • Planning Agent: Constructs DAG task graphs with explicit parent-child dependency bindings.
  • Execution Agent & Specialist Agents (Research, Communication): Interface with simulated digital provider portals.
  • Verification Agent: Performs independent state verification against system databases.
• AI Core: Directly interfaces with the Gemini API (gemini-2.5-flash) via developer API keys, avoiding any Google Cloud Platform overhead.
• Backend & Database: Python 3.11+, FastAPI, SQLAlchemy, and SQLite in WAL mode for local state persistence (ARCHITECTURE.md).
• Frontend: A React + TypeScript + Vite Command Center with dynamic visual DAG rendering and real-time approval queues.
• Mock Digital World: A dedicated mock server (mock_world/server.py) simulating real-world broadband, utility, and logistics APIs.

Challenges we ran into

1. Deterministic Safety in Non-Deterministic Agents: Large Language Models tend to execute actions eagerly. Enforcing strict application-level risk boundaries to prevent agents from auto-booking paid services required decoupling decision logic into a deterministic Python Risk Engine.
2. State Recovery & Paused Workflows: When an agent pauses mid-workflow to await user approval (which could take minutes or days), maintaining full state context across database persistence without losing agent step integrity was tricky.
3. Handling Asynchronous Real-World Failures: Simulating vendor cancellations mid-execution required building an event bus (Event Monitor) capable of waking up the Orchestrator agent to execute replanning loops smoothly.
4. Strict Zero-Cloud Enforcement: Ensuring zero GCP runtime dependencies while using Google ADK required strict environment sandboxing and custom validation scripts (scripts/check_no_gcp.py).

Accomplishments that we're proud of

• True Agentic Resilience: Demonstrating live autonomous replanning when a vendor emergency cancellation occurs—finding replacement movers and proposing a new plan without user prompting.
• Zero GCP Footprint: Running full Google ADK multi-agent workflows locally on standard consumer devices using standard Gemini API keys.
• Fact Confidence System: Extracting lease notice dates and policy conditions from unstructured text while assigning explicit confidence ratings to avoid hallucinated timelines.
• Verification over Assumptions: Implementing an agent that double-checks provider databases to confirm outcome states rather than blindly trusting function execution.

What we learned

• Verification is essential for agent reliability: Tools can report 200 OK, but the underlying operation might fail or linger in pending states. Verification agents bridge the gap between intent and reality.
• Human-in-the-Loop builds user trust: Users don't want black-box AI spending money. Clear approval cards showing costs, risks, and rationales make autonomous delegation feel safe and controllable.
• Local-first agent infrastructure is viable: Complex multi-agent workflows with local database persistence can deliver sub-second responsiveness without reliance on heavyweight cloud infrastructure.

What's next for LifeOps

• OAuth & Real-World Integrations: Expanding beyond mock provider APIs to real calendar tools, Gmail/Outlook APIs, utility portals, and WhatsApp notification hooks.
• Multi-Goal Concurrent Operations: Enabling LifeOps to manage multiple simultaneous life operations (e.g., wedding planning + apartment move + tax filings) with cross-workflow dependency tracking.
• Localized Mobile Command Center: Developing native iOS/Android companion apps to receive push notifications for high-risk approvals on the go.
• Edge LLM Support: Adding support for running small, quantized local open models (e.g., Gemma 2) alongside Gemini for offline privacy-sensitive document processing.

Built With

• adk
• antigravity
• gemini