Case Study

LLM for Beauty E-Commerce Store

Beauty online store
.Net, DynamoDB, AWS, Llama-2, Phi-2, Mistral-7b AI text generative models.
4 .Net Developers, 2 Automation QA-s, Business Analyst, Project Manager.

Challenge and Initial Analysis

Vaevi Technologies has proposed an improvement to an existing client, who was struggling with extracting information from lots of suppliers’ sources. These sources were a combination of PDF, CSV files and html-s. Our client required vast human effort to extract information from those sources into products’ description of a Beauty web store.

Vaevi has proposed a prototype, based on Llama-2 AI text generative model that could transfer PDF plain text into structured data format. The prototype has been reviewed by the customer’s engineering and management team and a contract has been signed.

Vaevi’s main delivery was targeted on the automation of the Web Store’s input process for products and reduction of human effort to do that.

Additionally, there was a requirement to compare several AI engines and pick up the most optimal solution for various languages: English, Ukrainian and Russian.

Solution Development

Vaevi’s team has utilized an external BA to discover all the possible flows and business rules of data extraction and transformation. As a result, 113 sources have been identified and documented. Basing on the business analysis result, our QA team has developed Integration tests and test data to evaluate effectiveness of the AI generative model.

Vaevi’s development team has made an evaluation and comparison of 3 AI text generative models:

Phi-2

Llama-2-7b-chat-hf

Mistral-7b-Instruct-v0.2

Llama-2 has been selected as the most effective solution.

The team has built a parsing and transformation module and connected it to the web store’s injection API.

The client needed to update the existing acceptance logic and API structure of the Web Store, to be able to integrate our solution’s output.

113 data sources
have been identified and documented

Implementation and Results

It took 6 months for the team to implement the solution and deliver it to the client’s Testing environment and qualify the results with the Integration tests.

The features included:

Automated check for updates on the supplier’s web origins,
transformation to the appropriate format of PDF, CSV and web sources on demand
injecting the acquired results into appropriate API for further human verification.
All the features have been demoed to the customer’s stakeholders, including CEO and the head of Human verification team.

After successful production launch, Vaevi has supported the solution for 3 more months with 1 Dev and 1 QA role.

Additional enhancements to this solution have been purchased by our customer as a separate contract and included:

Logging and data analysis of all the transformations of records to increase accuracy.
Integrated verification and scoring of injected records.

As a result of Vaevi’s solution:

our client was able to automatically inject supplier’s data into its Web Store with 69% less human effort involved*.
It simplified the work of data operators and reduced a number of data mistakes in the system.
Injection of the records and updates of the product’s details have become more often, resulting in reduced time-to-market.**

Let us guide you through the project lifecycle and work out the most fitting solution

Conclusion

This project exemplifies Vaevi Technologies’ abilities to comply with the most up-to-date tech demands of our clients and proactively propose improvements to their business.

*- Analysis has been made using the average human involvement time per record compared to the same metric in the first month after Prod launch. As KT factor was still in place, stable delivery result may be even better.

** – Relative subjective evaluation, based on the client’s management feedback.

Drag