Ganji Kumar

Generative AI (Nov’23 - Present): Developed a RAG application using GPT3.5, for technicians to query technical manuals. Used Opensearch to store chunk embeddings and deployed the application using Gradio. Extended it with a reranker for retriver and implemented a semantic cache to reduce API cost. Finetuned Llama2 7B Instruct Model for Logs summarization. Performed PEFT using Lora and updated weights of 0.3% parameters for better response. Implemented Slack Agent to receive summary. AI For Connected Living (Oct’22 - June’23): Developed audio signature classifier using YAMNet to classify audio from camera and notify Xfinity customers. Built IR images classifier usign XGboost model with image features (blob size, skewness, entropy, intensity statistics, SSIM). Implemented a object state detector using MobileNetV2 and CLIP embedding for various doorways. (POC). Primary for Object Detection Model in production. Captured 500 errors using shadow deployment and resolved them. Media Analytics (Jul’22 - Oct’22): In Celebrity Recognition model, Finetuned facial parameters and Improved clustering accuracy by adding functionality to upload new faces to elasticsearch. AIOps Initiatives (June’23 - Nov’23): Implemented kubeflow pipelines for automated training and deployment for deepant and prophet models for anomaly detection of metrics. Performed scaling and performance testing on the endpoint. Implemented multiprocessing for scheduling parallel training of 100’s of models in kubeflow.

Implemented influencer segmentation using k-means clustering model on youtube data to identify potential influencers with better channel engagement. Implemented youtube data scraping tool.

Electives - Machine Learning, Neural networks, CV, NLP, Information retrieval, Discrete Math., Linear Algebra, Statistics, Computational Finance, etc.