LeetCode tells you what to solve. It never answers the question that actually decides your interview: what should I practice next?

That is a decision problem, not a content problem — so Learn AI is built like every XingAI product: a decision system, not a chatbot. Paste an interview question (or a screenshot of one), and it answers with a pattern, the interviewer's intent, your gaps, and the next best question — company-aware.

Master patterns, not questions

The domain model is a knowledge graph: PatternFamily → Pattern → Question, with your LearningHistory, PatternMastery, and MistakeRecord layered on top. "Two Sum" matters only as evidence about how you handle hash-map patterns. Readiness is a score computed from pattern coverage, weakness reduction, and consistency — not a count of solved problems.

One LLM call, nine engines

The obvious architecture is one LLM call per concern: classify, extract intent, list prerequisites, map the pattern, explain, generate replay steps, find similar questions, recommend. Nine calls per question.

Learn AI makes one. A single analysis call returns one JSON document; the engines are parsers and post-processors over it. Only the engines that need the database — similar questions, company-aware recommendation, readiness — do their own work, against SQLite, with zero extra LLM calls (ADR-002).

You keep the modularity (each engine independently testable as a pure function) without the 9× cost, 9× latency, or nine ways for the answers to disagree with each other.

The worker standard, honestly deviated from

XingAI's global standard says: logic in a worker, API reads cache, frontend renders cache. That works when input is enumerable — N stocks, M meal plans. An interview question pasted at 11pm is not enumerable. There is nothing to precompute.

So Learn AI runs the LLM in the request path — guarded by three cache layers:

Layer	Key	What it saves
OCR	sha256(image)	vision call on re-uploaded screenshots
Analysis	hash(text + image)	the full analyze call on repeat questions
Similar questions	category:subcategory:pattern	per-pattern regeneration (serve 5 from 20 cached)

Cache hit → zero LLM involvement. Miss → exactly one call, cached for everyone after. Cost scales with distinct questions, not traffic — and in interview prep, everyone asks the same questions. The standard's goal ("no LLM call per page load") is met; the mechanism differs, and the deviation is recorded as an ADR instead of drifting silently.

What's next: grading our own recommendations

Learn AI just adopted the cross-product Decision Ledger (ADR-003) — every "practice this next" becomes a ledger row. Unlike most products, we can actually measure action_taken: if you analyze a question matching the recommended pattern within 7 days, the recommendation was followed; the next mastery update tells us whether it helped.

An interview-prep tool that grades its own advice against your outcomes — that is the difference between a decision system and a content library.