Aung Kaung Khant

A domain this wide strangles itself with circular imports if you let it. I grouped the backend by business process and kept the groups talking through an event/listener layer instead of importing one another.

• Payments the way people here actually pay — bank transfer, mobile wallet, QR and cash on delivery, with uploaded slips, an accountant approval stamp and deposit-plus-balance preorders.
• Query work with teeth — N+1 pulled out of listings and dashboard reports; pagination and transactions standardized everywhere.
• RBAC, audit logging, JWT + OTP with layered rate limiting, and Excel/PDF exports.
• Led a four-developer team — pull request reviews and the architecture documentation.

Nothing slow belongs on the request path. Broadcast notifications and payment-timeout jobs that auto-cancel unpaid orders run as BullMQ workers; Redis carries the cache and the short-lived redeem tokens. The API answers, the queue does the waiting.

• One domain, end to end — shops, menus, orders, riders, delivery zones, subscriptions, coupons, reviews and reporting.
• Realtime order tracking over WebSockets, with Firebase push so the customer hears about it with the app closed.
• S3 image pipeline with Sharp compression on upload.

The RubexPay gateway has no webhook. A guest who tips and closes the app would strand the payment mid-flight — so I built a scheduled reconciliation sweep that settles anything the customer never waited around to confirm.

• All three tiers, solo — the Laravel API, the Next.js admin dashboard, and the mobile PWA event guests actually hold.
• Gateway end to end — order creation, authorize/capture, returns and refunds.
• RSA-encrypted client-to-API payloads.
• Money people can check — income breakdowns with PDF and Excel payout reports.

Burmese retrieval was failing. Translating the question into English for the retrieval step only — while still answering in the language it was asked in — took top-result accuracy from 1/4 to 4/4 on a measured document probe.

• Hybrid search keeps identifiers findable — dense vectors plus BM25, with a cross-encoder reranker on top.
• 2.5s instead of 31s to first citations, once answers streamed over SSE.
• A 20-minute PDF embed that finishes — background ingestion with client polling instead of a request that times out.