Business Problem The client is a mid-size FMCG company, with a revenue of ~$150 M, 10 years old, having traditional IT services powering their application suite. Being only a 250-member team, with 70% of the employees concentrated in the Sales, R&D and manufacturing.
Supporting 55+ Product lines, with hundreds of variants, a meagre team of 10 category managers were having a hard time managing the e-commerce channel. On top of it, they had the distributor catalogs to manage, with latest product variants listing, constant updates on offers & prices. Ltd. budgets did not allow additional hiring, at the same time, manual listing & updation of the products was resulting a slowdown in go lives, inventory updates & eventually a hit on the sales, across the distributor network & e-comm channels. The client was looking to automate the products listing process, with high accuracy. But the challenge was the time to go live & availability of the images to train the model. Our Data Scientists & ML Engineers worked with their Category Management team to understand the listing process and then sat down with the R&D team to understand the various parameters used to build these products & how the competitive market is studied.
Solution
Post analysis, the team identified the initial set of 5 product lines, with the maximum variants & the most no of quality images, to do an initial pilot on automating this process. The team then leveraged Stable Diffusion, to understand the uploaded product image(s) & generate a compelling, yet well thought through, product description & the product USP. In addition, the system also scrapped the web to identify the competitive products & suggest the price range for listing the product, to ensure higher sales
An exquisite product image was meticulously uploaded and stored in AWS S3 bucket. We harnessed the power of the cutting-edge Hugging Face BLIP-2 Large Language Model (LLM) in Python to craft compelling text content based on the visual representation of the product. We leveraged Langchain to employ the image description as a prompt, which is then seamlessly integrated into the LLAMA-2-7B Model. This innovative synergy ensures that we obtain precisely the outcomes we desire. Our process included web scraping to identify similar products, driven by the text generation from the image achieved in Step 2. This iterative approach allowed us to furnish comprehensive results, including estimated price ranges, detailed product descriptions, and the unique selling propositions (USPs) of the products. To provide convenient access to these capabilities, we have harnessed FastAPI to create a suite of diverse and user-friendly APIs. The entire application is seamlessly deployed within an EKS Cluster, ensuring optimal performance and scalability.”
Architect
Outcome
- 72% automation achieved for products listing process, for the pilot product lines, covering 60+ variants.
- The system generated the target price point range for product listing, thereby increasing the conversion rates on the e-commerce channel.
- The process automation reduced the product go live time by more than 70%
