The Challenge
Headspace had a vast library of meditations, exercises, and wellness content—but connecting users to the right content at the right moment was a persistent problem. Generic recommendations led to drop-off; users needed something that understood where they were in the moment and could guide them toward what would actually help.
The vision was Ebb: an empathetic AI companion that could have genuine conversations with users, understand their current state, and connect them to relevant Headspace content through natural dialogue rather than static menus.
My Role
I led a team of six ML engineers and three fullstack engineers, driving Ebb’s technical strategy and execution across multiple workstreams:
Redesigned the Content Recommendation System: Headspace’s existing recommendations were based on simple heuristics and collaborative filtering. I rebuilt the system to incorporate conversational context—what users share in dialogue with Ebb directly informs which meditations, sleep content, or exercises get surfaced. This meant the recommendations feel personally relevant rather than algorithmic.
Built a Motivational Interviewing Agent: Rather than giving advice or pushing content, Ebb uses motivational interviewing techniques to help users explore their own thoughts and arrive at their own insights. I designed and implemented the conversational architecture that enables this—open-ended reflections, affirmations, and a dialogue flow that respects user autonomy. The agent guides without prescribing. Every feature decision balanced technical feasibility with clinical appropriateness through close partnership with clinical product managers.
Designed and Patented a Large Action Model (LAM): Invented a novel system for suggesting clinical skills and tasks based on conversational context—predating public LAM disclosures by major companies. This became the backbone of how Ebb connects conversation to action.
Built a Generative AI Wellness Journal: Designed, built, and safety-tested a hybrid agent functioning as a wellness journal, combining generative AI with structured therapeutic techniques. A/B testing showed a 34% increase in conversion to paid subscribers.
Optimized AI-Generated Smart Responses: Engineered the system that drafts responses for care team members, reducing clinical team edits per message by 67% and saving approximately 5 minutes per half-hour session (A/B tested).
Built a Custom Annotation Pipeline: Designed and built an internal annotation interface for LLM preference data, saving $125K/year in external vendor costs. Reduced annotation time by 75% while increasing annotator productivity 10x.
Mentored 4 MLEs from entry level to staff level, building the team’s capabilities alongside the product.
Technical Approach
The system architecture reflects the complexity of building AI for mental health at scale:
- Context-aware recommendation engine that weighs conversational signals alongside historical preferences
- Dialogue management system implementing motivational interviewing principles (OARS: Open questions, Affirmations, Reflections, Summaries)
- Safety layer with escalation paths to crisis resources when needed
- Personalization that remembers—Ebb picks up where previous conversations left off
Results
Ebb launched to Headspace members in the US, UK, Canada, and Australia:
- 2025 Webby Award for Best Use of AI/ML (People’s Voice Winner)
- 33% increase in engagement metrics (subscriptions, app opens, content starts) verified through causal analysis
- Users can speak or type, with Ebb responding conversationally in either mode