This article was created for the purpose of participating in the #AllThingsAgentic Hackathon by Devpost and Google.
Designing and building an app aimed at a young audience is a huge challenge—and that is exactly what I had to do when creating Ember.
In reality, Ember grew out of a genuine concern for my own son. I felt he was using his phone too much and spending excessive time immersed in video games, which was damaging his self-esteem. I wanted to create Ember to foster his personal growth through writing—an inherently introspective tool—and through face-to-face activities with others, such as sports, music, and art.
But how do you leverage the same Artificial Intelligence that typically fosters digital addiction to build a tool aimed at the exact opposite—promoting well-being, real-world connection, and mindful usage?
Algorithms vs. Real-World Action
The most common way to measure technology-driven applications is through metrics that track user attention and engagement: average session duration, daily active users (DAU), churn rate, and uninstalls.
However, for Ember, success is defined when the user closes the screen to engage in physical activities or write in their smart journal. AI should not optimize for mass consumption, but rather for trigger efficiency. Ember's core effectiveness metrics are evaluated as follows:
- Time-to-Value (TTV): The time it takes for a user to open the app, receive an AI suggestion (a writing prompt or workout routine), and initiate real-world action. Lower TTV indicates a more effective trigger.
- Conversion to Offline Action (COA): The percentage of AI recommendations that translate into a logged physical or creative action.
- Task Completion Rate (TCR): The proportion of sessions that end with a fulfilled goal (e.g., writing 500 words or training for 30 minutes) rather than passive browsing.
- Inverse Retention Metrics: Evaluating whether the user returns to the app not out of feed addiction, but because the AI successfully reinforced a real-life habit.
Agentic Architecture & Google Cloud Stack
Ember's backend is built on an autonomous, event-driven multi-agent architecture hosted on Google Cloud Platform. It utilizes the required Google AI framework stack to ensure low latency, scalability, and seamless background intelligence:
- Brain & Reasoning Engine (Gemini on Vertex AI): Gemini 3.5 Pro drives the complex reasoning required to analyze user writing logs, emotional cues, and long-term goal progression. Gemini 3.5 Flash handles real-time contextual evaluation with minimal latency to maintain an optimal Time-to-Value (TTV).
- On-Device Private Agent (Gemma 4): Runs locally on the user's device to process sensitive journal entries, private text logs, and personal reflections. By keeping intimate data on-device, Gemma guarantees total user privacy while extracting non-sensitive contextual features (like sentiment or activity intent) before communicating with the cloud.
- Agentic Framework (Google Agent Development Kit / Genkit): Orchestrates agent autonomy, planning, and tool execution. The framework allows agents to independently evaluate whether an intervention is needed based on contextual signals rather than waiting for direct user prompts.
- Asynchronous Background Processing (Cloud Run & Cloud Pub/Sub): Operates non-intrusive background tasks. Cloud Pub/Sub ingests real-time telemetry (sedentary time, app closure events, accelerometer data), triggering microservices hosted on Cloud Run that execute agent workflows even when the client app is closed.
- Long-Term Memory & Context Store (Firestore): Functions as the agent's persistent memory. Firestore stores user habit histories, baseline energy logs, and conversion rates, allowing Gemini to personalize future activity prompts without violating user privacy.
Multi-Agent Workflow Breakdown
- Context & Well-being Agent (Cloud Run + Pub/Sub): Evaluates incoming telemetry against Firestore baselines to determine the ideal window for intervention.
- Prompt Creation Agent (Genkit + Vertex AI): Generates structured offline action triggers tailored to the user's personal interests (sports, art, journaling).
- Evaluation & Feedback Agent: Monitors post-trigger telemetry to calculate the Over-engagement Penalty and update Ground Truth labels in Firestore.
Data Model Comparison
The following chart illustrates how user time is distributed during a typical session across traditional engagement models versus Ember's agentic model:
AI for Real-World Impact, Not Clicks
- Input Features: The AI evaluates contextual variables, including time of day, self-reported energy levels, local weather, and physical inactivity.
- Ground Truth Labels: A recommendation is classified as successful only if the user locks their device immediately afterward and sensors (accelerometer or health APIs) confirm movement, or if the editor records active writing bursts.
- Over-engagement Penalty: The optimization algorithm penalizes long sessions where the user merely navigates menus without initiating offline action.
Core Pillars
- Pillar 1: AI as a Mindful Companion: Processes reflective input to encourage real-world activities without introducing infinite scrolling mechanics.
- Pillar 2: Bridging Digital & Physical Worlds: Incentivizes real-world participation through healthy, non-addictive gamification.
Algorithmic Shift: Redefining Reward Functions
To build a truly agentic well-being platform, we must fundamentally overhaul the core optimization objective:
-
Traditional Apps:
Reward = Screen Time + Clicks -
Ember:
Reward = Offline Action Success / Screen Time Consumed
In Ember's model, the less screen time required to trigger an offline activity, the higher the model's reward score.
The Business Dilemma: Retention vs. Utility
Designing an app that encourages users to put their phones down raises a fundamental question: How does a platform survive when its primary goal is to minimize screen time?
Traditional digital platforms rely on ad revenue, forcing algorithms to maximize screen time and passive consumption. Ember rejects this model entirely. As a Social Impact Tech initiative, Ember shifts the financial model away from the individual user toward the institutions dedicated to their growth—universities, educational networks, libraries, and youth well-being organizations.
Through B2B institutional licensing and community impact grants, Ember secures sustainable funding without ever monetizing user attention or selling private data. The platform's value is not measured in ad impressions, but in human flourishing and real-world habit formation.
Conclusion: Building for Human Growth
As engineers shaping the next generation of Artificial Intelligence, we hold the ethical responsibility to define what our algorithms optimize for. We can build models that exploit human psychology for ad impressions, or we can build agentic systems that empower users to reclaim their lives, build real-world habits, and foster personal growth.
Ember is my answer to that choice. By prioritizing real-world action over screen engagement, Ember is the perfect definition of a technology-based project with social impact—using AI not to capture human attention, but to restore it.
Top comments (0)