Case study · Our own product
What we build when the client is us
TracLane is the largest thing we have built, and we built all of it: backend, web app, mobile app, integrations and deployment. It is how we show range without asking a client for permission.
Said plainly up front: TracLane is a product we own and are still building. It is not a client engagement, and we are not claiming a customer base for it. What it demonstrates is engineering range and integration depth.
The problem it solves
Pakistani e-commerce runs on cash on delivery, which means an order is not revenue until a courier hands over a parcel and collects money, and a large share never convert. Sellers manage this across spreadsheets, three courier portals and a phone.
TracLane pulls orders from Shopify and WooCommerce, routes them across warehouses, verifies cash-on-delivery orders by automated phone call, books and tracks couriers, and settles the money.
Scale
| Backend | Django 5.2, Django REST Framework, PostgreSQL, Celery and Redis, on AWS with continuous deployment |
|---|---|
| Web | Vue 3 and TypeScript |
| Mobile | React Native and Expo, Android and iOS |
| Integrations | Two e-commerce platforms, three couriers, Meta Marketing and Conversions APIs, WhatsApp Business, automated telephony, subscription billing |
The integration surface is the hard part. Three couriers with incompatible city-code databases, each with its own failure modes, all behind webhooks with retries and request logging.
Two pieces worth a technical look
Server-side ad tracking that survives the death of cookies
Orders are matched back to the advertisement that produced them, and real outcomes are sent to the ad platform rather than clicks. When a cash-on-delivery order is confirmed, delivered or returned, that event goes back with hashed customer data for match quality and a unique event identifier so it is not double-counted against the browser pixel. The optimiser then learns from what actually happened, not from checkouts that were never paid for.
A model router, not a chatbot
Four model providers sit behind one interface, including a self-hosted option. Queries are classified and routed on cost, speed or quality, with automatic fallback when a provider fails. The advertising recommendation engine is a deliberate hybrid: deterministic rules produce most of the value and the model supplies the explanation.
That second choice is the one worth discussing. Knowing which parts of a problem should not be handed to a model is most of the skill in shipping AI that behaves predictably. A rule you can read is worth more than a prompt you have to trust.
Have a model that needs to behave?
We build AI systems that are checkable: grounded answers, reproducible output, safety rules in code, and tests that gate every deploy. Tell us what is breaking and we will tell you whether we can fix it.
Start a conversationOr email info@codezbit.io