Part of the Amazon Retail Business Section, working on a business problem to reduce missing search attributes and to increase the discoverability of the products in the Amazon catalogue. I am working with Large Language Models (LLM) to predict the attribute value. Also building multi-lingual models in French, German and Italian.* Experimented with Finetuning of Mistral-7B using QLORA for the Irrelevant content removal problem yielding results better then DL solution.* Developed an irrelevant product detection model using S-BERT architecture with MSE loss function w.r.t. searched query and performed A/B testing which proved less irrelevant results in UK and AU. * Fine-tuned LLM's (Falcon-40B) using QLORA/PEFT techniques to improve model performance. After fine-tuning the model give better results from 82.5% to 91.2% better results (human audits). * Working with LLM's to reduce the defect rate of the catalogue and increase discoverability of products. Reduced the defect rate of catalog ~10% for prime products (2022) and 3.6% for all buyable products (2023). * Have built/fine-tuned multiple BERT, RoBERTa, T5 models to improve catalog defects in english/non-english marketplaces.

* Worked on MMM Automation setup to build regression models. Tweaked the contribution of marketing variable (TV, Radio, advertisement) more than 70% in regression models. The control variables contribution was mapped less than 30%. Different APL (Adstock-Power-Lag) combinations were used to tweak the regression models.* Implemented feature importance module in setup, that helped to identify important features, by backward elimination, forward selection, random forest feature selection etc.* Build an end-to-end data processing pipeline for complaints and faults use-case. Build procedures using T-SQL to add incremental data on monthly/quarterly basis in the final table.

* Build a multiclass-classification (8 classes) models using TensorFlow framework to classify the complaints from customer emails for an automated email intake setup. Accuracy = 92.3%* Build a Named Entity Recognition (NER) model using spacy library to map the customer data to relevant attributes and fill a final form, data mapping was using a docker images - Docanno. Deployed on client side as the model accuracy = 89.2% and it replaced a regex-based model in production.* The project was deployed on client machines and it reduced the manual force by 30%, with increase in customer success rate (rating) from 3.5 to 4.2.