Sparsh Rawal

• Built a RAG-Based conversational AI trained on Trustpilot Product Reviews. • Reduced CX issue detection time and generated customer review summaries using Anthropic’s Claude3 Sonnet. • Performed Market Segmentation(RFM) on 6 million customer records using unsupervised methods on AWS Sagemaker. • Visualized Demographic segments on UK map using Kepler.GL and Quicksight. • Created postcode-wise sentiment heatmaps from AO reviews using AWS services via prompt-engineering in AWS Bedrock (Claude 3.5). • Delivered comparative sentiment analysis of 300,000+ reviews across Currys, AO, and Marks Electricals using Claude 3.5 on AWS Bedrock. • Developed and labeled 5,000+ delivery images using SageMaker Labeling Jobs; trained a custom Rekognition model to identify faulty washing machine connections, reducing QA effort by 3 hours weekly. • Forecasted washing machine sales using AWS Forecast; achieved MAPE of 6.3% on the test set. • Created a GenAI-based SSIS documentation tool to automate audit processes. • Simulated pricing leader-follower scenarios using Monte Carlo simulations in SageMaker for 10 competitors and 4000+ products; insights integrated into pricing strategy. • Created multi-agent system using Anthropic Agentic AI framework. • Drafted technical documentation for R&D tax claims; contributed to a successful tax benefit of £50K+ based on innovation cases. • Worked in a cross-functional Agile team of 8 using Azure DevOps Boards for sprint planning, SageMaker for model development, and QuickSight for dashboards. Tools: AWS SageMaker, Bedrock, Transcribe, Forecast, Rekognition, QuickSight, Kepler.GL, Python, Pandas, Scikit-learn, Azure DevOps.

• The data used for the LLM was already in text format and was converted into chunks, each chunk containing 800 characters with an overlap of 100 characters. • Each chunk was converted into word embeddings using OpenAI’s text embedder and upserted to a Pinecone vector Database. Vector DBs are a key component for the RAG system due to their efficiency. • Designed the architectural pipeline for the chatbot, i.e. convert the user input into vector embeddings using the text-ada-2 embedder and query the pinecone database for retrieving the top 4 most relevant chunks based on the user input which was in tern used by the LLM. • Performed prompt engineering to create various prompt templates to be used based on the user query and ensure that the generated response complies with the privacy policies and industry standards • Worked on an experimental feature that enables the chatbot to interact with an user using voice-to-text and text-to-voice. • Enabled the Chatbot to accept an audio input from the user, this audio was converted into text using the SpeechRecognition python Library, this text was the input query to our chatbot. Also the generated text was converted into an audio output using the google TTS engine.

• Preprocessed the CVs by converting them into text files and removing unwanted extra information which was not relevant for matching the candidates with the JDs, such as personal information, educational information etc. • Converted the JDs and preprocessed CVs into text embeddings which is essential for clustering • Created 3 groups of JDs based on the required years of experience which were fresher, professional and senior. • Used the K-means clustering algorithm on each of the groups to cluster similar JDs inside each group. • Since we already know how many years of experience a candidate has, the CV embeddings were used to find the closest JD cluster within the respective group; the metric used was cosine similarity. • The top 5 closest JDs were extracted from the obtained cluster as the most relevant and potential match for the corresponding CV. This methodology ensures that the candidates are matched with the best JDs based on their CV. • Conducted experiments with LSTM, RNN, and Logistic Regression classifiers to identify the optimal model for Language Identification. Performance evaluation encompassed various metrics, including Accuracy, Precision, Recall, and F1-Score. • Extended the LSTM model to tackle Low Resource Languages, for which no training data was available. This included testing the model in a zero-shot setting for identifying low resource languages. • The entire development, training, and testing of the model were executed in Python, leveraging a range of libraries such as pandas, NumPy, matplotlib, and TensorFlow.

• Trained an ARIMA model on the weekly order data for the last 5 years, the model captures both the trend and seasonal patterns in the order history of the business customer. • Evaluated the model performance on metrics such as Mean Absolute error, RMSE and MAPE. • Integrated the model with the existing Amdocs OMS to capture order data directly from the OMS Database. • Wrote scheduling job scripts which load new data to the model weekly to ensure real-time data predictions. • Prepared documentation for the entire Model pipeline to ensure easy knowledge transfer and uninterrupted model output. • Trained an ARIMA model on the weekly order data for the last 5 years, the model captures both the trend and seasonal patterns in the order history of the business customer. • Evaluated the model performance on metrics such as Mean Absolute error, RMSE and MAPE. • Integrated the model with the existing Amdocs OMS to capture order data directly from the OMS Database. • Wrote scheduling job scripts which load new data to the model weekly to ensure real-time data predictions. • Prepared documentation for the entire Model pipeline to ensure easy knowledge transfer and uninterrupted model output.

• Employed data preprocessing techniques to clean and normalize the data, used one-hot encoding to convert the categorical features into numeric features. • Used SMOTE to handle class imbalance present in the dataset. Split data into Training, Development and Test set. • Trained a Deep Neural Classifier on the training set and fine tuned the hyperparameters on the development set. Used F1 Score as the evaluation metric and evaluated the trained model on the test set. • Integrated the model to the Comcast CRM system to enable customer retention techniques for the risk customers. • Created a model inference pipeline to classify a new customer. • Additionally, identified the most important features contributing to the model classification performance, to enable the marketing team to tailor customer retention strategies. • Created documentation for the model training and testing pipeline to enable easy knowledge transfer