# LIFEOPS — AI Chief of Staff
## Inspiration
We all live fragmented lives. Your calendar doesn't know your bank balance. Your fitness tracker doesn't know about your deadline tomorrow. Your email doesn't know you're already overwhelmed. Every morning, millions of people wake up and manually piece together their own lives across 12+ disconnected apps, making critical decisions with incomplete information, missing patterns that were completely predictable, and burning 2+ hours a day just on the mental overhead of context switching.
We didn't want to build another chatbot. We wanted to build what we actually needed — something that understood our whole life at once, not just the slice that lived in one app. The moment we asked ourselves *"what if one agent had access to everything and could reason across all of it simultaneously?"* — LIFEOPS was born.
The inspiration was deeply personal. We've all been in that stress spiral — too many meetings, bad sleep, stress eating, skipped workouts, overspending — and realized in hindsight that every signal was there days before it hit. The data existed. It just lived in five different apps that never talked to each other. We wanted an agent that would see that spiral coming before we felt it, and do something about it.
---
## What It Does
LIFEOPS is an AI Chief of Staff that connects every app in your life into a single intelligent agent that understands your patterns, reasons across every domain simultaneously, takes real action on your behalf, and gets measurably smarter about you every single week.
### It Connects Everything
Using Fivetran, LIFEOPS automatically syncs data from Gmail, Google Calendar, Google Fit, Plaid, Amazon, DoorDash, Spotify, Notion, GitHub, Strava, Uber, and more into a central BigQuery data warehouse. You connect once. It syncs forever.
### It Understands Everything
All synced data is enriched, categorized, and indexed into Elasticsearch with vector embeddings, making it instantly searchable across every domain simultaneously. MongoDB Atlas stores a permanent, evolving memory of the user — preferences, patterns, past decisions, relationship details, and behavioral triggers that grow richer every day.
### It Reasons Across Everything
Powered by Gemini 2.5 Pro and orchestrated via Google ADK, LIFEOPS runs specialized subagents for calendar intelligence, financial monitoring, health tracking, social relationships, shopping decisions, and habit coaching. When you ask a question, every subagent queries simultaneously and the orchestrator synthesizes a complete, cross-domain answer.
### It Acts Proactively
LIFEOPS doesn't wait to be asked. It monitors your data 24/7 on Cloud Run, sending a personalized morning briefing every day, detecting stress spiral patterns before they hit, flagging budget overruns before they happen, and reminding you about relationships that need attention. When it detects something, it proposes specific actions and executes them with a single approval tap.
### It Improves Itself
Every recommendation LIFEOPS makes is fully traced in Arize Phoenix via OpenInference instrumentation. Every week, LIFEOPS uses the Phoenix MCP server to query its own trace data, evaluate which recommendations were followed and which produced good outcomes, identify its own blind spots, and update its behavior accordingly. It is the only personal AI that gets measurably better at helping you the longer you use it.
---
## How We Built It
We built LIFEOPS on a five-layer architecture where each layer has a distinct responsibility and the layers compose into something far more powerful than any individual component.
### Layer 1 — Data Collection (Fivetran)
We set up Fivetran connectors for every major life data source. Fivetran handles the complexity of OAuth, incremental syncs, schema changes, and API rate limits so we never have to. All raw data lands in BigQuery on a continuous sync schedule.
### Layer 2 — Data Enrichment (BigQuery)
We wrote transformation jobs in BigQuery that clean, categorize, and enrich the raw data. Transactions get ML-classified into spending categories. Emails get summarized and action items extracted. Health metrics get combined into composite scores. Behavioral patterns get detected using SQL window functions across rolling time periods. The output is a clean, unified view of the user's life ready for intelligent querying.
### Layer 3 — Search and Retrieval (Elasticsearch)
Enriched data from BigQuery syncs into Elasticsearch Serverless with vector embeddings generated using Google's `text-embedding-004` model. We designed eight specialized indices:
- `life_events`
- `financial_intelligence`
- `health_timeline`
- `communication_intel`
- `pattern_library`
- `agent_insights`
- `habits_data`
- `preferences`
Each index is optimized for its specific query patterns. We built custom ES|QL tools that power specific agent capabilities like budget tracking, stress detection, relationship health checks, and behavioral pattern matching. These tools are exposed to the agent via the Elastic MCP server so Gemini can call them at runtime without custom API code.
### Layer 4 — Memory and Reasoning (MongoDB Atlas + Google ADK + Gemini 2.5 Pro)
MongoDB Atlas serves as the permanent memory layer — storing user profiles, relationship details, past decisions and their outcomes, discovered patterns, and weekly retrospectives. We built seven specialized subagents using Google ADK:
- Calendar Intelligence
- Financial Intelligence
- Health Monitoring
- Social Relationships
- Shopping Decisions
- Habits and Goals
- Travel Planning
Each subagent has its own system prompt, tool access, and domain expertise. A master orchestrator coordinates them, fans out queries in parallel, and synthesizes responses using Gemini 2.5 Pro's long-context reasoning capabilities.
### Layer 5 — Observability and Self-Improvement (Arize Phoenix)
We instrumented the entire agent runtime with OpenInference using the `google-adk` instrumentor, sending every trace — every tool call, every LLM decision, every subagent interaction — to Arize Phoenix Cloud. We configured the Phoenix MCP server so LIFEOPS can query its own traces at runtime as a tool. We built an LLM-as-Judge evaluation pipeline using Phoenix Evals that scores every recommendation on:
- Contextual fit
- Actionability
- Follow-through
- Outcome quality
Every Sunday, the self-improvement engine runs — querying its own traces, analyzing evaluation scores, identifying failure patterns, and updating its own system behavior. This is the closed loop that makes LIFEOPS genuinely improve over time.
### Infrastructure
The entire system runs on:
- **Google Cloud Run** — 24/7 proactive monitoring
- **Cloud Build** — CI/CD pipeline
- **FastAPI** — backend powering the web chat interface and webhook handlers for real-time data events
---
## Challenges We Ran Into
### Cross-Domain Reasoning Coherence
The hardest technical problem was making multi-domain synthesis feel natural rather than mechanical. When a user asks a simple question like *"help me plan my week,"* the agent needs to pull from calendar, finance, health, and social simultaneously and synthesize it into a single coherent response — not a list of five separate reports. Getting Gemini to truly synthesize rather than just concatenate required significant prompt engineering and careful orchestration of how subagent outputs were structured and weighted.
### Data Pipeline Reliability
Fivetran connectors have different sync frequencies, different schema formats, and different data quality characteristics for every source. Building the BigQuery transformation layer to handle missing data, schema drift, duplicate records, and late-arriving data without breaking downstream Elasticsearch queries was more complex than anticipated. We had to build defensive transformation logic for every connector.
### Vector Search Quality for Personal Data
General-purpose embeddings don't always capture the nuance of personal life data. The phrase *"stress eating"* needs to semantically match *"ordered DoorDash at 11pm three days in a row"* — that's a domain-specific semantic relationship that generic embeddings handle poorly. We spent significant time on embedding strategy, chunking approaches, and hybrid search tuning to get retrieval quality high enough for the agent to reason correctly.
### The Self-Improvement Loop Integrity
Making the agent query its own traces and actually change its behavior based on what it finds — without hallucinating patterns that don't exist or over-correcting on noise — required careful design of the evaluation pipeline, minimum sample sizes before patterns are considered reliable, and guardrails on how aggressively the system prompt could be updated in a single week.
### Privacy Architecture
LIFEOPS handles extremely sensitive personal data. Designing a system that is genuinely useful — meaning it needs to understand real patterns in real data — while ensuring that data never leaves the user's own infrastructure, that email content is summarized but not stored verbatim, and that financial details are aggregated before being passed to LLM prompts required careful thought at every layer of the architecture.
### Proactive vs. Intrusive Balance
Building a proactive monitoring system that alerts users before problems happen is the core value proposition. But an agent that notifies you too frequently, or that flags things you don't care about, becomes noise that users turn off. Getting the threshold calibration right — when to alert, what urgency level, which channel — required extensive tuning and ultimately became something the self-improvement loop helps handle automatically by learning which alert types the user responds to.
---
## Accomplishments That We're Proud Of
### The Self-Improvement Loop Actually Works
Watching LIFEOPS query its own Arize Phoenix traces, identify that its health recommendations had a 34% follow-through rate while financial recommendations had 85%, diagnose that the health advice was too vague, update its own approach, and then see follow-through climb to 61% the next week — that was the moment this stopped feeling like a demo and started feeling like something real. Building an agent that genuinely improves its behavior based on its own operational data is technically difficult and we got it to work cleanly.
### True Cross-Domain Reasoning
When a user types *"I feel overwhelmed"* and LIFEOPS comes back with a response that simultaneously accounts for their meeting load, sleep data, budget status, upcoming social commitments, and a pattern it detected 7 times before — and proposes five specific actions across different systems — that is not something any existing personal AI tool can do. We are proud that this works the way we imagined it.
### The Full Data Pipeline End to End
Getting Fivetran syncing real data → BigQuery transforming and enriching it → Elasticsearch indexing with vector embeddings → the agent querying via MCP in real time — as a complete, working pipeline across four different technologies — is a real engineering accomplishment. Each integration alone is non-trivial. All four working together reliably is something we're proud of.
### The Pattern Detection Accuracy
The behavioral pattern engine — detecting that a user always stress-eats on Thursdays after heavy meeting days, or that they consistently overspend on Friday evenings, or that their productivity peaks on Tuesday mornings — and detecting these patterns from real behavioral data with enough confidence to act on them proactively, is genuinely useful in a way that feels different from typical AI features.
### Proactive Intervention That Feels Helpful, Not Creepy
Getting the proactive alert system to feel like a caring chief of staff rather than surveillance software required getting a lot of details right — tone, timing, specificity, the approval flow before any action is taken. We're proud that in user testing, the most common reaction to the proactive alerts was *"how did it know?"* rather than *"this is uncomfortable."*
---
## What We Learned
### Data Is the Product
The quality of the Fivetran pipeline and the BigQuery transformations determine the ceiling of what the agent can ever achieve. A brilliant LLM with bad data produces confidently wrong answers. We learned to invest in data quality first and agent sophistication second.
### MCP Is Genuinely Transformative for Agent Architecture
Before this project, MCP felt like a nice abstraction. After building LIFEOPS, we understand it as a fundamental architectural pattern. The ability to expose Elasticsearch queries, MongoDB operations, and Phoenix trace data as tools that any MCP-compatible agent can call — without writing custom integration code for each — changes how you think about composing agent capabilities.
### Self-Improving Agents Require Humility Engineering
The hardest part of the self-improvement loop was not the technical implementation — it was preventing the agent from over-learning. An agent that updates its behavior aggressively based on small samples will oscillate rather than improve. We learned that good self-improvement requires minimum confidence thresholds, rolling averages rather than point-in-time evaluations, and conservative update magnitudes. Humility has to be engineered in.
### Proactive Is the Right Model
Every user we showed LIFEOPS to had the same reaction — they didn't want to have to ask it things. They wanted it to tell them things. The proactive monitoring and alerting system resonated more deeply than any conversational feature we built. The future of personal AI is not better chat. It is better anticipation.
### Cross-Domain Synthesis Is Gemini's Strongest Suit
We tried several approaches to multi-domain reasoning and found that Gemini 2.5 Pro's long-context capability was uniquely suited to holding the outputs of multiple subagents simultaneously and synthesizing them into a coherent response. The quality of cross-domain synthesis improved dramatically as we gave it more complete context rather than trying to compress or summarize subagent outputs.
### Observability Is Not Optional for Agents
Without Arize Phoenix tracing every decision, LIFEOPS would be a black box that either works or doesn't. With full observability, we can see exactly why a recommendation was made, which tools were called, what data influenced the decision, and where things went wrong. This is not a nice-to-have for production agents — it is the difference between an agent you can trust and one you can only hope works correctly.
---
## What's Next for LifeOps
### Native Mobile Application
The proactive alert system is most powerful when it lives on your phone. A native iOS and Android app with background processing, rich push notifications, and quick-action responses would dramatically increase the value of the proactive monitoring capability. The morning brief, pattern interrupt alerts, and one-tap approvals all deserve a first-class mobile experience.
### Expanded Action Capabilities
Right now LIFEOPS proposes actions and executes them with approval. The next step is deeper action integration — direct calendar event modification, actual bill payment, real grocery orders placed, gift purchases completed, meeting declines sent. The goal is reducing the gap between *"LIFEOPS suggests"* and *"LIFEOPS handles"* to the point where the user's only job is approval.
### Richer Relationship Intelligence
The social and relationships layer has enormous untapped potential. Pulling in LinkedIn data, shared calendar availability, communication sentiment trends, and gift purchase history to build a truly comprehensive relationship management system — one that feels like having a personal assistant who has been with you for years — is a major roadmap item.
### Multi-User Household Mode
Most of the problems LIFEOPS solves are not individual problems — they are household problems. Shared budgets, shared calendars, coordinated meal planning, joint savings goals. A household mode where multiple users' data is combined into a shared intelligence layer while preserving individual privacy would make LIFEOPS relevant to families, couples, and shared living situations.
### Vertical Specializations
The core LIFEOPS architecture can be specialized for specific contexts:
- **LIFEOPS for Founders** — integrates startup metrics alongside personal life
- **LIFEOPS for Athletes** — weights health and recovery data more heavily
- **LIFEOPS for Parents** — incorporates children's schedules and activities
The foundation is built. The specializations are a matter of domain-specific tooling and prompt tuning.
### Deeper Self-Improvement Capabilities
The current self-improvement loop updates the agent's behavior weekly. The next evolution is real-time adaptation — the agent updating its approach mid-conversation based on signals that a recommendation isn't landing, and running continuous A/B experiments on its own prompt variants using Phoenix's experiment framework to find what works best for each individual user.
### Enterprise and Team Version
The same architecture that manages one person's life can manage a team's operations. A LIFEOPS for teams would integrate project management tools, team calendars, budget tracking, and communication platforms to give team leads the same cross-domain intelligence at the organizational level — predicting team burnout, flagging budget overruns, and coordinating complex multi-person workflows automatically.
### The Long-Term Vision
An agent that knows you so well, has been with you so long, and has accumulated so much verified knowledge about what helps you specifically — that it becomes the one system you would never want to give up. Not because it's locked in, but because the depth of understanding it has built about you over months and years is irreplaceable.
**That is what we are building toward.**
Log in or sign up for Devpost to join the conversation.