The client deals with the problem of voucher creation (multiple fields in the voucher) out of a large pool of crawled content data.

Current solution incorporates semi-automatic voucher creation using recent advancements in NLP and availability of pretrained models for language generation.

Due to the fact that the voucher generation latency was not suitable for on demand generation, the main goal was to speed up the creation process, at the same time freeing up more resources.

Based on the problem analysis and the needs of the client, the suggested solution includes fine tuning of separate models for each individual field in the voucher. The suggested execution environment was decoupled from the business process and horizontally scalable (including GPU support).

Also, the model’s time execution complexity was reduced by using onnx runtime instead of a native execution runtime and by float number representation precision decreasing. 

According to the results (in terms of time complexity) obtained on the initial benchmark, the  evaluation of the proposed models and the suggested horizontal architecture the voucher generation latency is appropriate for real-time on-demand generation on high loads.

Project team:

Gjorgji Madjarov

Vladimir Trajkovikj

Igor Mishkovski

Stefan Andonov