Sai Dheeraj Gummadi

• AI Software Engineer, Working on Designing and Implementing Inhouse MultiModal RAG System for Internal Employees to understand different products based on their documentations(Text, Image, Video formats).

• As a Deep Learning Engineer at Brane, I Designed and Developed the pipeline for freeflow text based No Code Low Code Platform using Natural Language Processing techniques and LLMs which can build the complete solution from the given set of sentences in natural language
• Synthesised data of workflows in different domains using ChatGPT, LLama2, and Mistral for training T5/BART/GPT models for coreference resolution, paraphrasing, splitting complex sentences, span extraction
• Implemented knowledge distillation to create distilled versions of LLMs and deployed the models as RESTful APIs using Docker and Ray Serve.
• Training BART for paraphrasing and Splitting the complex sentences.
• Fine-Tuned GPT2 (1.5 billion parameter) for next task prediction using LORA and Bitsandbytes on multiple GPUs (Tesla V100) using model and data parallelization. A virtual assistant tailored for solution developers, aimed at enhancing their productivity.
• Deployed the model with vLLM Inference Engine and achieved low-latency (0.06 sec) and high-throughput (71 rps) inferences by optimizing real-time interactions.
• Prototyping RAG based approach for next task prediction using LlamaIndex for getting less hallucinated responses from LLMs.

• As Data Scientist at HighRadius, Designed and Developed an end-to-end pipeline for a Financial Document Scanner product using CNN, YOLO, and OCR for extracting the required tabular regions like balance sheets, and income statements from Financial Statements and analyzing the customer’s credit risk.
• Trained and Improved the 6-layered unconventional CNN using Explainable AI tools like Lime and GradCam
• Achieved an overall automation rate of 80% and Reduced 40% of the time spent on manual analysis of financial documents with 100–200 pages.
• Improved the latency of the Generic Parser product which is one point solution for parsing Claims/Invoices/Remittences/ POD’s using NLP transformer-based quantized LayoutLM token classification model and unsupervised algorithms with an overall accuracy of 90% and achieved automation rate of 75%.
• Worked on POC for classifying the pages in input documents into Sales Invoice, Remittance, Check, and Claim for improving the accuracy of downstream parsing modules by Fine Tuning the LayoutLM Sequence Classification model for classifying the images based on both content and layout.
• Developed Payment date prediction models using regression models like Random Forest / XGBoost for FORTUNE 500 CPG companies to predict invoice payment date with an average +/-3 Day accuracy of 85%.

Master of Science in Data Science at the International University of Applied Sciences, Berlin, Germany. I studied masters parallely while working as data scientist at HighRadius.

B.Tech at the SRM University, AP with a major in Electronics and Communication Engineering. I Published 8 conference papers on deep learning published in IEEE, Springer, Scopus databases and Patented a pipeline for disease diagnosis under Anirban Ghosh.