ClassFlow Live Tutoring Marketplace Architecture

Operational coordinators manually reviewed student requests, teacher subject strengths, timezone differences, and calendar slots, creating up to 4-hour booking turnaround times and frequent scheduling conflicts.

Designed a real-time heuristic scoring pipeline evaluating 3-tier weighting (timezone offsets, historical completion ratings, and active instructor workload) resolving matches in under 100ms.

Algorithmic assignment requires continuous calibration of teacher load ceilings to prevent superstar burnout, but reduced match latency by 99%.

Peak registration bursts created race conditions where multiple students attempted to confirm bookings for the same teacher calendar window at the exact same sub-second interval.

Implemented a distributed lock pattern using Redis Redlock with a 45-second lease time and atomic lease release on checkout completion, ensuring strict mutual exclusion across nodes.

Introduced a hard operational dependency on Redis cluster availability, backed by automated fallback to database row-level locking during failovers.

Polling HTTP endpoints for classroom events (student check-in, instructor late warnings, connection drops, session completion) generated massive request volume and a 10-15s state lag.

Built a real-time event pipeline over FastAPI WebSockets paired with Redis Pub/Sub, streaming classroom state transitions and automated grace-period timers to all parties with sub-second latency.

Stateful socket connections required dedicated sticky-session management and reconnect jitter logic on client apps.

Instructors were paid via manual spreadsheet calculations at week-end, resulting in high administrative labor, human calculation discrepancies, and delayed disbursements.

Designed an event-driven webhook pipeline consuming verified session completion events, computing commission splits with PostgreSQL Serializable transactions, and dispatching Stripe Transfers idempotently.

Required cryptographic webhook signature validation and dead-letter queue replay mechanisms, but reduced operational payout workload to zero.