Abhishek Sharma Email & Phone Number

Kota, Rj, In

Delhi, India

Deployed LlaMA3 and LangChain-based Qustion Answering engine locally, enabling faster queries on 20 GB+ local documents without internet connectivity.Enhanced QnA engine responsiveness to <3 secs and scale to 5 TB+ documents using Milvus DB and gte-large-en-v1.5 achieving more efficient similarity searches and improved query accuracy migrating from Chroma DB and hkunlp-instructor.Refined prompt styles and hyper-parameters (top-p, top-k, temperature etc) to achieve… Show more Deployed LlaMA3 and LangChain-based Qustion Answering engine locally, enabling faster queries on 20 GB+ local documents without internet connectivity.Enhanced QnA engine responsiveness to <3 secs and scale to 5 TB+ documents using Milvus DB and gte-large-en-v1.5 achieving more efficient similarity searches and improved query accuracy migrating from Chroma DB and hkunlp-instructor.Refined prompt styles and hyper-parameters (top-p, top-k, temperature etc) to achieve classification accuracy of 90%+ and reduced hallucination to <1%.Fine tuned LLMs using Parameter Efficient Fine-tuning (PEFT) method LoRA for different report types improving robustness of the model.Designed a Flask API-powered UI for document summarizing, and LLM training via feedback by user, capable of scaling upto 500% (5x) user engagement.Automated SOP and Report generation with LlaMA3 achieving 90% manual effortreduction using zero shot learning eliminating need of training data. Show less

Mumbai, Maharashtra, India

SKUs Forecasting: Developed XGBoost models for critical SKU demand forecasting, reducing inventory waste by 30% through hourly predictions of critical SKUs.Refactored and deployed ML Inference Pipeline using Docker and AWS EC2; set up CICD lifecycle with automated retraining for data drift, boosting inference speed by 75% and maintaining forecast accuracy.Order and Rider Prediction: Trained XGBoost models to predict orders and rider requirements for specific time slots… Show more SKUs Forecasting: Developed XGBoost models for critical SKU demand forecasting, reducing inventory waste by 30% through hourly predictions of critical SKUs.Refactored and deployed ML Inference Pipeline using Docker and AWS EC2; set up CICD lifecycle with automated retraining for data drift, boosting inference speed by 75% and maintaining forecast accuracy.Order and Rider Prediction: Trained XGBoost models to predict orders and rider requirements for specific time slots, achieving 3x improvement in forecasting accuracy with 21-day simulations.Constructed User Address Correction algorithm using Pytorch, BERT, and Reverse Geocoding API; decreased Escalation KPI, Where is the Order, by 40%.Developed and Implemented an RFM-based analytics framework, executing targeted campaigns through data-driven segmentation and achieving 4% monthly retention increase and 55% ROI uplift.Designed BI Dashboards on AWS Quicksight to monitor and drive Business KPIs and wrote 20+ SQL queries. Show less

Mumbai, Maharashtra, India

Image Captioning using Convolution Neural Network as encoder and Recurrent Neural Network as decoder.Optical Character Recognition and used YOLO for object detection task.Automated reply system for Email using Python.