Selected work
One year. Mostly backend. Built where the failure modes mattered.
messages.sql
postgres
-- Keyset cursor pagination · index seek, not table scan
SELECT m.*, u.name, u.avatar
FROM messages m
JOIN users u ON u.id = m.sender_id
WHERE m.conversation_id = $1
AND m.created_at < $2 -- cursor: last seen ts
ORDER BY m.created_at DESC
LIMIT $3;End-to-end messaging — DMs, group channels, presence, read receipts. Keyset cursor pagination replaced an O(n) offset scan with an O(log n) index seek. A Redis cache layer covers the hot path so PostgreSQL barely sees it.
Event-driven analytics: thousands of LMS events per day flow through Lambda into ClickHouse with YYYYMM partitioning and composite sort keys. A regex classifier maps raw events into nine categories, configurable without redeploy.
webhook.ts
typescript
// Webhook is the source of truth — client confirmation is never trusted
const event = stripe.webhooks.constructEvent(
req.body,
req.headers["stripe-signature"],
process.env.STRIPE_WEBHOOK_SECRET,
);
if (event.type === "payment_intent.succeeded") {
await Order.markPaid(event.data.object.metadata.orderId);
}Full-stack event platform with Stripe payments, JWT auth, and role-based access control. Order state lives in the webhook handler — never trusted to the client confirmation — so a closed browser doesn't lose a paid order.
Only one case study is shipped today. The others sit on a queue — deep dives in the order I have something new to say.