← All cases← All cases
● LIVE PRODUCTCASE / 03
AI Telegram product · 2026
Leela AI
A self-discovery game where deterministic rules protect the canon and AI creates personal meaning.
Open @lel_ai_bot ↗PRODUCT SYSTEM● LIVE
01 / PROBLEM
The challenge
Every move needed a personal interpretation connected to the player’s intention and previous answers. Sending the entire conversation to an LLM each turn would cost about $0.50 per game and make the core rules impossible to reproduce reliably.
02 / RESPONSE
The solution
We separated responsibilities: a deterministic engine owns the 68-cell Johari board, dice, arrows, snakes and state transitions; a Smart-AI layer owns interpretation. A rolling summary compresses the player’s history to roughly 900 tokens per turn, while fallbacks and a crisis-signal safety layer keep the experience resilient.
03 / SYSTEM
Product architecture
The core operating flow connecting the product idea, code and a measurable outcome.PRODUCT SYSTEM● LIVE
04 / EVIDENCE
Product in use
Real screens and working product states.
05 / MODULES
Inside the system
06 / OUTPUT
Personal AI depth with production-grade economics
Verified outcomeslower estimated LLM cost per game
unit tests covering the game engine
simulated games completed successfully
“Code owns the rules. AI owns the meaning.”
Architecture principle — Leela AI
Next case / 04Unf*cking Horoscope↗