Aztec Scanner — Computer vision library · proprietary

Reads the Aztec code on Polish vehicle registration certificates from a phone photo, so insurance chatbots get exact car data without typing.

Role: Author

Client: Insurance assistance chatbots, Poland

Team: Solo

Period: 2018–2019

Car-insurance assistance chatbots in Poland need exact vehicle data — make, model, VIN, registration number — and asking drivers to type it is slow and error-prone. Every Polish registration certificate carries an Aztec 2D code holding that data, but decoding it from real phone photos, which are skewed, unevenly lit and low-resolution, is far harder than reading a clean scan. The bots needed a scanner that copes with such photos and plugs into their PHP back end.

I built the decoder as a Python pipeline with OpenCV and NumPy, compiling the heavy image-processing modules with Cython for speed. Contour detection locates the symbol; grid detection, enhancement and reading turn pixels into bits; then the mode message and data layers are parsed. Reed–Solomon error correction recovers damaged codewords, and the certificate payload is decompressed and decoded into fields. A PHP wrapper with GD runs the pipeline, maps fields to the bot’s data model and turns failures into friendly retry messages.

Locates the Aztec symbol in skewed, noisy phone photos

Grid detection, enhancement and bit reading compiled with Cython

Reed–Solomon error correction for damaged codewords

Decodes the certificate payload into mapped vehicle fields

PHP wrapper with error handling and user-friendly retry messages

Used in insurance assistance chatbots; 30 commits over five months; closed source

Python, Cython, OpenCV, NumPy, PHP, GD, Node.js, Reed–Solomon