We partnered with a Berlin health tech startup to design and build an AI-first mental wellness product — from zero to a working MVP in 6 weeks that secured €1.2M in seed funding.
The Challenge
Two founders — a clinical psychologist and a software engineer — had a compelling thesis: daily micro-reflections guided by AI could meaningfully improve mental wellness for young professionals. They had domain expertise, early user research with 50 potential users, and a clear vision.
What they lacked was AI architecture experience. Their first attempt — a ChatGPT wrapper with custom prompts — produced inconsistent responses, forgot context between sessions, and felt generic. Users described it as “talking to a search engine about my feelings.”
They needed:
- An AI system that remembered — each user’s history, preferences, emotional patterns, and progress over time
- Consistent personality — a warm, professional tone that felt like a trusted guide, not a chatbot
- Clinical safety — appropriate boundaries, crisis detection, and escalation protocols
- Speed — they had 8 weeks of runway before their next investor meeting
Our Approach
We applied the Ophelin Method with a startup-optimized cadence — compressing the typical timeline while maintaining architectural rigor.
Observe (Week 1)
Instead of lengthy discovery, we ran an intensive 3-day workshop:
- Day 1: Reviewed the founders’ user research, mapped 12 core user journeys, and identified the single most valuable interaction pattern
- Day 2: Analyzed 200 sample conversations from their prototype to understand where the AI failed and why
- Day 3: Defined the MVP scope — one core feature (daily guided reflection) done exceptionally well, with everything else deferred
The critical insight: users did not want therapy. They wanted a structured moment of clarity in their day. This reframing shaped every design decision.
Architect (Week 2)
We designed a lightweight but extensible architecture:
- Context Engine — Stores and retrieves user history, emotional patterns, and session summaries. Each conversation builds on the last.
- Personality Layer — A carefully crafted system prompt architecture that maintains consistent tone, adapts to user communication style, and respects clinical boundaries.
- Safety Framework — Keyword detection, sentiment analysis, and escalation protocols that route crisis situations to human resources.
We chose a modular architecture specifically so the startup could extend it post-funding without rewriting the core.
Refine (Week 3-5)
Three weeks of build-test-refine cycles:
- Week 3: Core reflection flow operational. 10 beta testers onboarded.
- Week 4: Context engine refined based on real conversations. Personality tuning based on user feedback (“too formal” → adjusted to warm-professional).
- Week 5: Safety framework validated with clinical advisor. Edge cases addressed. Onboarding flow added.
Emerge (Week 6)
Beta launch with 50 users from the founders’ waitlist. We set up monitoring, analytics, and feedback collection. The system was live, learning, and improving from day one.
The Solution
The final MVP consisted of three core components:
-
Daily Reflection Flow — A 5-minute guided reflection that adapts its questions based on the user’s mood, recent entries, and long-term patterns. Not a questionnaire — a conversation.
-
Memory & Progress — The AI remembers themes, tracks emotional patterns over time, and references previous sessions naturally. Users reported feeling “actually understood” — the primary differentiator from generic wellness apps.
-
Safety Net — Invisible to most users, but critical. The system detects concerning patterns and provides appropriate resources without breaking the conversational flow.
The tech stack was intentionally lean: Cloudflare Workers for the API, a React Native frontend (built by the founders’ engineer), and OpenAI’s API with a custom prompt architecture that we designed.
Key Takeaways
-
Scope ruthlessly. The founders’ original vision included journaling, meditation, community features, and therapist matching. We shipped one feature — daily reflection — and it was enough to prove the thesis and raise funding.
-
AI personality is a product decision, not a technical one. We spent more time on tone, boundaries, and conversation design than on model selection. The “how it feels” matters more than the “how it works” for consumer AI products.
-
Architecture for extension, not perfection. The MVP architecture was designed to be extended, not rewritten. Six months later, the startup added journaling and progress tracking without touching the core reflection engine.
-
Speed and quality are not opposites. Six weeks was enough because we made clear decisions early. The Observe phase — especially the “one feature” decision — eliminated weeks of scope creep.