One-time Deep Sync imports the user's entire LeetCode solve history via authenticated GraphQL — every accepted problem, not just recent submissions. Solved slugs are indexed against 1,500+ curated sheet problems and auto-checked off across Blind 75, NeetCode 150, Striver SDE/A2Z, Love Babbar 450, and company sets. Analytics surfaces solve breakdowns, topic mastery, language stats, a 12-week activity heatmap, streaks, XP, and contest ratings — all fan-out in parallel from PostgreSQL via Prisma. API responses ship with Cache-Control stale-while-revalidate headers; the client layer uses a localStorage SWR cache so pages paint instantly on reload.

Eight curated and company-wise sheets ship as structured JSON corpora, indexed into a flat slug map at boot. When Deep Sync runs, a single pass marks matching problems solved on every sheet the user follows — no manual checkboxes. Focus sheets pin to the dashboard; progress bars, difficulty breakdowns, and section-level completion update in real time. Guests can browse the full sheet library with sample progress to see the product before signing up.

The AI coach pulls a rich context snapshot — LeetCode stats, weakest topics, streak state, active quests, and sheet progress — then grounds replies with lightweight retrieval over 1,500+ sheet problems, 21 interview pattern families, and system-design notes. Lexical scoring over a pre-built corpus keeps answers citable without hallucinating; the chunk format is embedding-ready for a future pgvector upgrade. Personality modes range from supportive to savage; proactive nudges fire when streaks break or solve counts stall.

Global and friends-only leaderboards rank by XP, level, streak, solve counts, or LeetCode contest rating. Server-side unstable_cache pre-fetches the top 200 candidates and re-sorts per axis without extra DB round-trips — 60s revalidation keeps rankings fresh under load. Multiplayer coding rooms generate unique 6-char codes, track live solve progress, and power mock-contest study sessions. XP (Easy 15 / Medium 40 / Hard 120), coins, badges, and adaptive daily/weekly/revision quests auto-verify against real LeetCode solves.

Sentry is wired across Node.js server, Edge runtime, and client via @sentry/nextjs instrumentation — 10% trace sampling, graceful no-op when DSN is unset, and onRequestError forwarding for every server render and API route failure. PostHog runs dual-track: client-side SPA pageviews with Supabase user identification, plus server-side capture (posthog-node) for signup, onboarding, subscription, and referral funnels so backend state matches analytics dashboards. Both integrations no-op in dev when keys are missing.

Caching is layered at every tier so read-heavy endpoints never hammer the DB: Next.js unstable_cache (60s revalidation) for leaderboard and public profiles, HTTP Cache-Control with s-maxage + stale-while-revalidate on analytics/dashboard APIs, in-memory memoization for the 1,500+ problem corpus, and a namespaced localStorage SWR cache on the client (wiped on logout). Edge/CDN caching on Vercel absorbs anonymous guest-preview traffic on /sheets, /patterns, and /analytics. API hardening: Supabase session auth on every write path, middleware splits guest-preview routes from strict auth-required ones (/rooms, /chat, /settings), and LeetCode GraphQL calls are isolated with no-store fetch to avoid stale credential leaks. PostgreSQL + Prisma with parallel Promise.all fan-out and encrypted session cookies — built to scale from hundreds to millions of users on serverless without UX degradation.