In-Store Customer Behaviour Tracking Prototype

Problem

A web agency wanted to build a product for physical stores. The idea was to find out how customers move through a shop and what they do there, in the same way a website shows how visitors behave online. Nobody had built this before in the company, so it started with a prototype that had to prove the idea.

Approach

• I designed the technical side of the system and chose which features went into the first version.
• I built the prototype with a small team and wrote the key parts myself.
• The tracking device was a simple Android phone on a shopping cart, with an app I wrote that ran in the background.
• We tried infrared and NFC to find the position of a cart. Neither worked well, so we moved to Bluetooth beacons.
• I built the backend in C# with my team.
• Later, the phones were replaced by small custom chips. I wrote the code that ran on them.

Architecture and design decisions

• The system had four parts: a device on the cart, beacons in the store, a backend and a nightly analysis.
• The device sent its position to the backend. The position came from the Bluetooth beacons placed around the shop.
• The backend collected the data during the day. At night it went through many gigabytes of it and turned it into statistics about customer behaviour.
• The backend ran on C#, and MongoDB held the tracking data.
• The device code was first an Android app and later C and C++ on the chips.

Reliability and performance

• I dropped infrared and NFC after they gave poor results and chose Bluetooth beacons instead.
• I ran the heavy analysis at night, so it did not slow down the system during store hours.

Outcomes

• A working prototype of a customer tracking system for physical stores.
• The prototype became the base of a new company, which took the system on as its main product.
• The system moved from Android phones to custom hardware without changing the idea behind it.

Known gaps

• I have no numbers on how many stores, carts or customers the prototype tracked.
• I do not say how accurate the positions were, or how the statistics were used by shops.
• I do not describe the exact size of the data or the way the nightly analysis worked.
• The page does not say how much of the later product still comes from my prototype.