MonetKit — cross-platform subscription platform
A subscription management platform in the spirit of RevenueCat: one integration that normalizes the full billing lifecycle across Apple App Store, Google Play, and Stripe — so product teams stop re-implementing store quirks and start shipping.
Apps selling subscriptions on iOS, Android, and web each speak a different billing dialect — different receipts, webhooks, renewal semantics, and edge cases. The goal: one platform that normalizes all three so a product team integrates once and trusts the state it reads.
- —Core backend & domain model
- —Public API + SDK API — versioned, idempotent
- —Webhook pipeline & delivery guarantees
- —Entitlement engine
- —Admin portal frontend
- 01Canonical event model — Apple, Google, and Stripe lifecycles mapped into one internal vocabulary before any business logic runs.
- 02Transactional Outbox — state change and outgoing event commit atomically; a delivery worker retries with backoff.
- 03Idempotency everywhere — provider retries and out-of-order events are safe by design, not by luck.
- 04Entitlements as derived state — recomputable from transaction history, which makes debugging and audits tractable.
OTTclouds — live streaming infrastructure
Built and optimized HLS/LL-HLS video delivery with a custom player and analytics — improving playback, making viewer experience measurable, and keeping live operations calm.
Live video punishes every weak link: an origin hiccup or a slow segment shows up instantly as buffering on someone's screen. The work: make delivery fast, observable, and resilient — and know the system's real ceiling before viewers find it.
- —HLS / LL-HLS pipeline setup & optimization
- —Custom video player integration + analytics events
- —CDN & cache strategy — CloudFront, Lambda@Edge
- —Stream failover & load testing with Locust
- 01LL-HLS only where latency pays — low-latency for interactive streams, standard HLS for reach; a per-stream trade-off, not a dogma.
- 02Failover at the edge — Lambda@Edge / CloudFront Functions handle origin routing and failover instead of client-side hacks.
- 03Player analytics as a first-class system — startup time, rebuffering, and bitrate switches feed operations, not just reports.
- 04Load test to the real ceiling — Locust scenarios against the delivery path before peak events, not after incidents.
Reliable billing webhook pipeline
An event-driven pipeline hardened against the reality of billing webhooks: duplicates, out-of-order delivery, and services that crash mid-processing — without ever corrupting subscription state.
Billing webhooks are hostile input: providers retry aggressively, deliver out of order, and a worker can die halfway through a state change. A naive handler corrupts subscription state in ways customers notice on their invoice.
- —Pipeline architecture end-to-end
- —Deduplication & idempotency layer
- —Transactional Outbox + delivery workers
- —Dead-letter queue & replay tooling
- —Audit trail of every state transition
- 01Persist raw first, process later — the verified payload is the source of truth; processing is replayable from it at any time.
- 02Dedupe on provider event id — plus idempotency keys on every side effect, so retries are no-ops.
- 03Backoff + DLQ, never silent drops — poison messages land in a dead-letter queue with tooling to inspect and replay.
- 04Audit every transition — support can always answer "why is this customer premium?" from the trail.
Backend observability platform
Centralized logging, metrics, and distributed tracing on the LGTM stack — one place to see, correlate, and explain backend behavior, so incidents get shorter and debugging gets boring.
Debugging across services with scattered logs meant slow, stressful incident response — every investigation started with "which box do I ssh into?". The goal: correlated signals and dashboards that answer questions instead of raising them.
- —OpenTelemetry instrumentation across services
- —Loki · Tempo · Mimir deployment
- —Grafana dashboards per audience
- —Alerting rules & on-call ergonomics
- 01Correlate by trace id — logs, traces, and metrics share ids, so you jump from an alert to the exact request.
- 02Dashboards per audience — on-call triage boards stay minimal; deep-debug boards carry the detail.
- 03Alert on symptoms, not causes — latency and error-rate SLOs page people; internals stay as diagnostics.
- 04Vendor-neutral instrumentation — OpenTelemetry keeps the door open to Datadog and friends without re-instrumenting.