You don't hand me a ticket. You hand me a problem.

Orders were punched into a form and retyped into a spreadsheet. Stock went out to everyone as another spreadsheet, so the warehouse, the sales floor and the accounts desk each held a different file. Every decision was made by someone reading a sheet that was already out of date.

Stop patching Excel and build the system underneath it. The cheap plan — lock the cells, share the file properly — left the real problem untouched, because a spreadsheet cannot tell you who changed a number or which copy is wrong.

Structured order entry, stock modelled down to colour and roll, nineteen standing reports, and an uploader that reconciles the old spreadsheets against the database rather than banning them.

Operators were retyping handwritten textile packing lists into the ERP by hand, every day. The obvious answer was OCR, and the obvious answer did not work: handwriting in mixed Hindi and English defeated every single-model pipeline we tried.

Run two vision models over the same document and cross-check their structured output. One model was half the price, but nobody would have trusted it enough to stop double-checking — and a tool people double-check has saved nobody any time.

Async ingestion on BullMQ, confidence surfaced per field, and a human review queue for the rows the models disagreed on.

We needed real role-based access across a hundred-plus routes. Buying it was faster on paper. The problem with buying it was that our access model would then live in someone else's schema, and every future module would have to ask their permission to exist.

Build it. A Module → SubModule → Action tree, multi-role aggregation, and a two-tier cache — in-memory Map in front of Redis on a five-minute TTL — so the check costs nothing on the hot path.

Denied-access audit logging, rate limiting and dynamic IP blocking, and a PermissionGuard component gating the UI against the same tree the API enforces.

Busy ERP was the accounting system the business actually ran on. A cutover weekend was the standard plan and the standard plan was unacceptable — if it went wrong on Saturday, nobody could invoice on Monday.

Never migrate. Fence the legacy system behind a bridge service and let both run at once, so the modern stack never talks to Busy directly and either side can be rolled back on its own.

Live stock and balance events over RabbitMQ with auto-reconnect and gzip decompression, two-way master-data sync across 8+ entity types, and CLI scripts with --dry-run before anything touched production.

The app took a minute to become usable. Operators had learned to open it and go make tea, which is the kind of workaround that tells you nobody expects the software to improve.

Measure before touching anything. It was images and bundle weight, not the database everyone blamed — so the fix was boring and the database work we had queued up got cancelled.

Reworked image pipeline and frontend bundles, then moved the platform to AWS on Docker with zero-downtime CI/CD, which tripled API throughput on the way.

Three services were each growing their own notification code, and each was about to grow its own retry bug. Channel routing was leaking into business logic everywhere.

Services emit one event and stop caring. Routing, templating, retries, opt-in and delivery tracking are somebody else's job — a dedicated service on its own queues.

Per-channel RabbitMQ queues, MD5 deduplication so a retry never double-sends, per-user channel opt-in, and delivery-status webhooks back to the source service.