Case study
The AI feature I removed
AI PRODUCT MANAGER · CLEARA · NOV 2025 – PRESENT
The problem
Cleara is an AI-guided learning product built around structured practice sessions — short, repeatable, and designed to be returned to. I started it to build something end to end myself rather than describe what someone else should build.
The decision
The first version generated each session dynamically with an LLM. It was the more impressive approach, and on paper it was the right one — every session personalized, nothing hard-coded.
In practice the output was unpredictable. Not wrong exactly, but different enough each time that you could not build a habit around it. The core loop of a learning product is repetition, and repetition needs something stable to repeat.
I replaced the generative core with deterministic logic, and kept the model only where it actually earned its place: short, specific responses inside a structure that does not move.
What I built
- Structured learner flows across multiple session lengths and use cases
- A feedback loop that surfaces whether someone is actually improving, rather than optimizing for time in the app
- Product architecture covering state, navigation, and handling for AI failure and edge cases
- Repeated scope cuts toward fewer decisions per screen
What it is, exactly
A working end-to-end prototype. No users, no revenue, not publicly deployed. It was built to learn, and the learning is the result: a clear position on where AI belongs in a learning product and where it gets in the way. I am happy to demo it live.
What I would do next
Put it in front of ten people and watch where they stall. I have opinions about the opening flow that I have not yet earned the right to hold.