Duolingo-style AI concepts learner
The assignment
Build an app that makes you learn AI concepts the Duolingo way. Wire it up with PostHog for analytics.
✳What Sakshi shipped
AI Levels is a card-based learning app organized into modules with a visible progress bar and locked/unlocked lesson states. Its first module covers how AI models predict rather than ‘know’ answers, the role of randomness and temperature, how prompt wording affects output, and context-window degradation — framed around real, specific points of confusion learners have about AI behavior.
✳Our read on this build
Sakshi shipped a deployed web app you can open. This read comes from the artifact and Sakshi's own account of it — no scores, just what the build does and does not show.
What it demonstrates
- The return loop is real: levels and bite-sized lessons — the mechanics that make a habit instead of a bookmark.
- The syllabus is the real vocabulary — prompting — not a tour of how AI is changing everything.
What is not shown
- No analytics wiring is visible in what is public, so how the app is really used stays unknown — including to its builder.
Take this awayTwo things travel from this build: short lessons beat comprehensive ones, and events defined before launch beat opinions formed after it.
✳Why this quest matters
Retention is a design problem, not a content problem. Wiring PostHog in from day one is the tell of a builder who ships to learn from real behaviour — rather than to admire their own screenshot.
✳What separated the strongest builds
A loop worth coming back to
Bite-sized lessons, instant feedback, and a reason to return tomorrow — streaks, XP, spaced review. A long article with a quiz at the end is not the Duolingo shape.
Concepts that are worth the streak
Tokens, embeddings, attention, hallucination, RAG — the vocabulary a PM needs in a real conversation, not 'what is AI' trivia.
Instrumented, not just launched
PostHog wired in and firing real events. An analytics claim nobody can verify tells a reader nothing about how the product is used.
No numeric scores are published here. These pages are written reads of what each builder shipped — the build is the point, not a mark out of a hundred.