Projects at Ericsson:End-to-End ML Pipelines: Spearheaded the development of machine learning pipelines for managed service networks, encompassing problem formulation, exploratory data analysis (EDA), feasibility analysis, modeling, feedback loops, and productization.Test Case Generation Pipeline: Developed a code generation pipeline utilizing Mixtral 8x7b, Code Llama, and Tree-sitter for Java and Erlang languages, streamlining feature request handling.Conversational Assistant (RAG-based): Built a Retrieval Augmented Generation (RAG) conversational assistant for finance and accounting directives, achieving 96% accuracy in the retrieval stage using models such as MPNet, MiniLM, and Llama2 7b.LLM Evaluation: Conducted comprehensive studies on the evaluation of Large Language Models (LLMs), leveraging tools like RAGAS, Deepchecks, and Arize to ensure robust performance metrics.KPI Throughput & Latency Prediction: Developed predictive models for throughput and latency degradation in telecom networks (MBNL, T-Mobile, Verizon) using time-series data and models like PySpark, XGBoost, and LSTM.ML Ops Framework: Conducted a feasibility study on deploying IoT-related machine learning use cases using Kubeflow and Google Kubernetes Engine.Label Reader: Built a scene text detection and recognition pipeline using PaddleOCR and Azure Kubernetes Services (AKS) to extract product information from images at telecom site locations.Smart Forest Monitoring: Designed and deployed an ML pipeline using MobileNet and YOLOv5 on edge devices (Raspberry Pi) to detect and count animals in forest environments via Balena Cloud.

Led and guided end-to-end development in content intelligence and user profiling initiatives:Doubt Resolution: Developed prediction capabilities to analyze academic images or text and retrieve the closest related question or concept.De-duplication Layer: Designed a parallel neural network architecture to detect and flag duplicate academic questions.Smart Tagging: Built prediction models to identify key attributes (subject, chapter, concept) related to academic questions.Concept Mastery: Created algorithms to predict skill or concept mastery for students using the platform.Academic Name Disambiguation: Implemented clustering techniques to map academic noun phrases to their respective concepts.Technologies used: Auto Encoders, Universal Sentence Encoders, Language Models, Xgboost, DeepCTR, BERT, UMLFit, Deep Knowledge Tracing

Worked on a range of projects focused on healthcare analytics, building scalable machine learning models to address various challenges in the medical domain:Sentiment Analysis: Developed a sentiment analysis model to classify patient reviews of healthcare providers (physicians, payers) with high accuracy.Technologies used: LSTM, F1 Score: 0.97Prediction Layer in Medical Technology Devices: Implemented a prediction layer to forecast market trends in medical devices, reducing manual effort by 80%.Technologies used: Random Forests, F1 Score: 0.92Data Standardization: Designed a data standardization framework for medical procedure and test names using K-Means++ clustering algorithm.Disease Prediction: Built a classification model to predict diseases for patients based on their therapy or treatment group.Technologies used: LightGBM, AUC: 0.9, F1 Score: 0.94Patient Readmission Rate Prediction: Developed a model to predict unplanned hospital readmissions using historical US healthcare claims data.Technologies used: Logistic Regression, AUC: 0.87, F1 Score: 0.91

Worked on implementing machine learning models to enhance HR optimization solutions by automating processes and improving decision-making efficiency:HR Optimization Platform: Integrated machine learning models into the HR platform to streamline candidate evaluation.Classification: Developed Naïve Bayes-based text classifiers to categorize resumes into domains such as IT, FMCG, Mining, Automobile, and Manufacturing.Performance: Achieved an F1-score of 0.88 on the test set.Clustering Algorithms: Designed and developed clustering approaches (DBSCAN and K-Means++) for analyzing a corpus of resumes, job descriptions, and company reports.Resume Parser: Built parsing algorithms using Naïve Bayes to extract personally identifiable information (PII) from resumes, achieving an F1-score of 0.95.Employee Joining Propensity: Developed a logistic classifier to predict the likelihood of a candidate joining a company, achieving an F1-score of 0.96 in offline datasets and 0.9 in production.Skill Repository: Created a semantic knowledge layer for skills in the IT industry using Neo4j graph database, enriched by word2vec embeddings.

Led the development and optimization of data-driven platforms and solutions:Batch Data Platform: Collaborated with the analytics team on feature extraction, model tuning, and production deployment of data models.Model Productization: Productized models using the PMML feature in Spark to streamline model deployment.MapReduce to Spark Migration: Converted the default MapReduce codebase to a more efficient Spark-based environment.Customer Targeting Models: Applied Logistic Regression and Random Forest to identify the right customers for product targeting.

Ad-hoc Search Tool: Developed a custom search tool over raw data using Elasticsearch to enhance data retrieval.PxOE Solutions: Designed and developed the client-side framework of the Px-OE platform. Migrated client-side storage to Cassandra and automated the deployment of Predictive Experience Solutions for service delivery teams.

Contributed to the design and development of key distributed systems and internal frameworks:MapReduce Framework: Played a crucial role in the design and development of the internal MapReduce framework, optimizing the processing of large-scale data.Distributed System: Led the conversion of a standalone system to a distributed scenario, improving scalability and system performance.Monitoring Tools: Developed and implemented tools for monitoring the application processes across internal node networks.