● Leading development of a DARPA-funded AI project which helps trade schools personalize education● Engineered serverless Python microservices leveraging RAG and fine-tuned LLMs, resulting in 60% improvement in content relevance for school curriculum● Deployed ETL data pipelines on AWS which loaded data into MongoDB (NoSQL) and sent database changes to a RabbitMQ queue using the CDC pattern● Deployed training pipeline on Qwak that fine tunes Mistral 7b LLM using LoRA, monitored via… Show more ● Leading development of a DARPA-funded AI project which helps trade schools personalize education● Engineered serverless Python microservices leveraging RAG and fine-tuned LLMs, resulting in 60% improvement in content relevance for school curriculum● Deployed ETL data pipelines on AWS which loaded data into MongoDB (NoSQL) and sent database changes to a RabbitMQ queue using the CDC pattern● Deployed training pipeline on Qwak that fine tunes Mistral 7b LLM using LoRA, monitored via CometML (MLOps platform) to save the best model to its registry● Developed inference pipeline with model quantization and deployed it as REST API using FastAPI, which reduced average response time from 500ms to 300ms, while handling 1,000 requests per min Show less

● Enhanced ML infrastructure and algorithms, which helped scale the startup's ARR from $1M to $7.5M● Built XGBoost model to improve accident severity prediction accuracy from 68% to 81% AUC by leveraging Pytorch to fine-tune BERT for industry-specific embeddings, thereby reducing SMB customer churn from 30% to 12%● Deployed anomaly detection system using ARIMA-based changepoint algorithms using a Docker container on AWS Lambda● Designed a dynamic pricing model for fleet driving… Show more ● Enhanced ML infrastructure and algorithms, which helped scale the startup's ARR from $1M to $7.5M● Built XGBoost model to improve accident severity prediction accuracy from 68% to 81% AUC by leveraging Pytorch to fine-tune BERT for industry-specific embeddings, thereby reducing SMB customer churn from 30% to 12%● Deployed anomaly detection system using ARIMA-based changepoint algorithms using a Docker container on AWS Lambda● Designed a dynamic pricing model for fleet driving risks, resulting in an average 20% reduction in commercial insurance premiums for clients while enabling precise risk pricing Show less

● Boosted certification exam pass rates by 15% through a Random Forest model that predicted student churn with 85% recall, enabling schools to tailor educational paths and intervene appropriately● Using Python (pandas) and Principal Component Analysis (PCA) to reduce input features from 11 to 4, analyzed exam performance of 50,000 students to develop feature intuition for the most dominant factors affecting churn rate● Increased user satisfaction scores by 30% in one quarter by… Show more ● Boosted certification exam pass rates by 15% through a Random Forest model that predicted student churn with 85% recall, enabling schools to tailor educational paths and intervene appropriately● Using Python (pandas) and Principal Component Analysis (PCA) to reduce input features from 11 to 4, analyzed exam performance of 50,000 students to develop feature intuition for the most dominant factors affecting churn rate● Increased user satisfaction scores by 30% in one quarter by developing a RoBERTa-based sentiment analysis tool which analyzed feedback from 50,000+ users, and deployed it as Docker container to generate weekly scores on satisfaction KPIs (AWS Quicksight) Show less

● Led ML initiatives for seismic image processing for Shell, BP, and ExxonMobil. Developed a CNN model using Keras achieving 97% accuracy in classifying seismic noise, advancing digital signal processing automation● Reduced QC team processing time by 90% (from 1 hour to 5 minutes) by deploying a Flask RestAPI on Google App Engine, dramatically enhancing seismic data processing efficiency