Summary

Overview

Work History

Education

Skills

Address

Websites, Portfolios and Profiles

Languages

Timeline

UMADAS RAVINDRAN

Summary

19 years of IT industry experience specializing in text analytics, machine learning, natural language processing, deep learning, computer vision, and generative AI. Currently part of AICoE team at EdgeVerve AI Next, focusing on rapid prototyping to extract insights from structured and unstructured data such as PDFs, invoices, and emails. Proficient in data analysis, addressing class distribution, data imbalance, and augmentation prior to AI implementation, while staying updated on market trends to deliver tailored customer solutions.

Overview

years of professional experience

Work History

Senior Product Architect

Edgeverve Systems Limited

Bangalore, India

05.2020 - Current

Developed a segmentation model using public datasets to detect damages in cars with a predefined taxonomy of parts and damage types, using the ensemble of DeTR & SAM3 model with promptable concept segmentation, which achieved an accuracy of more than 70% on real-world damaged cars.
Developed fine-tuned small language models(Phi-3.5-mini-instruct/Llama 3.2 3b) on a shipping/logistic domain dataset, with instruction fine-tuning, in an Nvidia GPU for information extraction task with PEFT(LoRA) technique which got integrated into the AI pipeline of the product achieving an accuracy of over 90% overall.
Developed the code for a RAG based pipeline with GPT 3.5/Llama 3 & re-ranker techniques which has been integrated into the AI platform. Recently extended it to incorporate multi modal RAG with GPT-4o which can perform document QA with respect to images/tables etc. in addition to text. Also added RAG Triad technique to evaluate the pipeline with respect to relevance scores.
Fine tuned a flan-t5-base pretrained model with PEFT(LoRA) with multitask training (i.e. classification and question answering) for a financial domain dataset and demonstrated how a single model can perform various tasks, acting as a replacement to an existing complex solution(with N models) pipeline for the domain client.
Developed an innovative pipeline for information extraction task using DeTR model(identifying objects such as tables, paragraph, headers etc from documents) and Table Transformer model(structure recognition) to extract tabular fields from pdf documents for a financial customer to start with, which helped clinch deals from other clients too.
Developed the algorithms to fine tune Roberta and Bert/Distilbert models for Intent classification/NER/Question Answering downstream tasks using tensorflow & pytorch framework for a banking client, which eventually got integrated into the AI offering as a flask based APIs(deployed as a docker image on prem in customer environment or on cloud) invoked by the React based product UI.
Fine tuned a sentence transformer model for sentence similarity task for a financial domain and proved domain specific fine-tuning to be better than the pretrained model with respect to accuracy for the tabular text of the domain.
Developed the code for fine-tuning LayoutLMv1 model for image classification task with financial domain data and integrated into the platform which performs with an accuracy of over 95%.
Evaluated BERTopic package for shipment/logistic domain client data as a first level filter for a email classification solution pipeline which unearthed potential classes/labels of the emails.
Developed a solution for fine-tuning a financial customer dataset in SQUAD V2 format for question answering task using Roberta model which performs with an accuracy of more than 90% on the domain dataset.
Developed a quick POC to demonstrate model serving with respect to Ray Serve, integrating the product's NLP, Vision, Gen AI models which has replaced the existing custom model serving framework in the product.

Product Technical Architect

Infosys Limited

Bangalore, India

07.2006 - 05.2020

Improvised the keyword extraction algorithm using OpenNLP and parts of speech logic and developed the clustering algorithm from scratch using Spark LDA and ULMFit for the ticket analysis tool which gives valuable insights on a ticket dump , with respect to root cause, similar tickets, important keywords etc ,built on top of Big Data technologies.
Developed the prototype and the code for doing intent classification with CNN architecture using Keras embedding layer, using tensorflow framework and then performing named entity recognition on the classified intent using spacy as a consolidated pipeline, which acted as a lightweight solution for various client's use cases.
Showcased POCs to run distributed training with respect to CNN for text classification in Nvidia GPU with "tensorflow mirrored strategy" technique in AKS(Azure Kubernetes Service) in 2017 when containers/orchestration etc were not as popular as in current times. Impact created was faster training with respect to these models using the full capacity of the infra.
Engineered the tool "iSee(Infosys Semantic Extraction Engine)" which used libraries like OpenNLP , parts of speech parser, NLP techniques etc to extract fields like "actors", "action" from textual data for a well known American publishing company, which acted as an additional features for training machine learning models alongside the traditional BOW(N Gram)/TF-IDF approach.
Developed the semantic search module using Apache Lucene, with a configurable Elasticsearch & SOLR module for iSee which fetches Top-K similar text from the storage for the given text with confidence scores.
Worked for the R & D division of Infosys(i.e. SETLabs) in "Text Analytics" team, during the earliest phase of my career, using NLP techniques like Lemmatization, BOW/N Grams, TF-IDF, Ontology(OWL), OHE etc with respect to use cases like Sentiment Analysis, Opinion Mining, Classification using SVM, NER using Stanford NLP for internal subunits on textual data in Java stack.

Education

Bachelor of Engineering - Information Technology

Raipur Institute of Technology

Raipur

07.2005

Skills

Machine Learning:

Regression
Random Forest

SVM

Naive Bayes

Decision Trees

KNN

Clustering

Computer Vision:
Object Detection
Image Classification

Image Segmentation

NLP:

Text Analytics
Information Extraction
BOW/TF-IDF/Lemmatization
Sentence Similarity
NLTK/Spacy

Classification

RNN

LSTM/GRU

Data Science:
Python 3x
NumPy
Pandas
Scikit-learn

Matplotlib

Seaborn

Jupyter

Deep Learning:
TensorFlow 2x

PyTorch 2x

CNN

Transformers

Containers/Orchestration:

Docker/Podman

K8S

Model Serving:

Ray Serve

TF Serving

Torch Serving

Tools/IDE/Framework:
PyCharm

Eclipse
Flask/FastAPI

Label Studio

Nvidia Nemo

Gen AI:
LangChain
Fine Tuning
AWS Bedrock
RAG

Prompt Engineering

FAISS/ChromaDB

Streamlit

LLM/SLM

Nvidia GPU(A100)

Cloud Platforms:
Microsoft Azure

Additional Softwares:
Databases: Postgres, MySQL
Version Control: GitHub/Git

Address

D201, Ramanuja Enclave, Kodipalya Road, Kengeri, BANGALORE, 560060