LlamaIndex - AI Engineer And Developer Advocate
◦ Developed Evaluation modules for RAG system, first to propose LLM as judge in the RAG space as an early contributor of LlamaIndex.
◦ Integrated multiple data loaders for efficient data ingestion in the OSS framework.
◦ Implemented advanced RAG research including GraphRAG, CorrectiveRAG, AdaptiveRAG, and Mixture Of Agents as LlamaPacks.
◦ Developed an Orielly Media course on Building RAG Applications with LlamaIndex.
◦ Integrated cutting-edge LLMs, embedding models, rerankers, and fine-tuning APIs from OpenAI, Anthropic, Cohere, and Mistral.
◦ Conducted analysis and benchmarking for retrieval and generation aspects of RAG systems to aid corporate efficiency.
◦ Provided crucial support for issues and PRs within the OSS framework.
◦ Enhanced retrieval performance in LlamaCloud by benchmarking complex queries with sub-query planning and metadata filtering.
◦ Supported clients such as Byte Dance, EY, NetApp India, Albus, Atomic Works, and Videoverse in developing efficient RAG systems.
AI Researcher - Part-time - Independent OSS Work
◦ Developed the BRAG series of Small Language Models (SLMs) optimized for RAG, significantly outperforming established models like Cohere’s Command R+, Qwen2, Llama3.1, and Llama3 Instruct models and closely matching GPT-4-Turbo, Nvidia’s ChatQA-1.5-8B. Models and blogs are available here and here.
1. Improved performance through strategic data mix and minimal dataset exploitation using LoRA and QLoRA.
2. Achieved cost-effective training with each model trained under 25.
3. Work done in collaboration with Pratik under maximalists.ai.
◦ Created Navarasa 2.0, a Gemma finetuned model for 15 Indian languages, featured in GoogleIO keynote 2024. Models and blogs are available here and here.
1. Managed end-to-end aspects of model creation including data preparation, modeling, and evaluation.
2. Positioned Navarasa 2.0 among the top-5 models for Indian languages as evaluated by Microsoft Research.
3. Work done in collaboration with Ramsri under TeluguLLMLabs.
Glance - Inmobi - Senior Machine Learning Engineer.
◦ Created and deployed the pipeline to reduce the time taken for publishing screen saver on TV by 92% from 1 hour to 5 minutes.
◦ Developed and deployed Glance TV Screen Saver product in automating the process of creating a wallpaper on TV for the latest news article. Efforts include:
1. Generating short headlines and sub-headlines from the article text.
2. Image search engine using CLIP Embeddings and Sentence Transformers.
3. Text stitching of generated headlines and images from image search.
◦ Developed and deployed a transformer-based (Peaguses, GPT-3) paraphrasing system for comment generation on Glance TV and Live to improve user engagement.
◦ Developed and deployed an automated poll generation framework using GPT-3 to improve the user experience and engagement on Glance TV thereby increasing the watch time of poll interacting users by 32.07 % and by reducing the effort needed by content and editorial teams by 30%. Blog - here.
◦ Developed and deployed a deep learning based DropoutNet recommendation model which showed an improvement in interaction rate and impressions seen by cold and sparse users.
◦ Developed and deployed an auto-encoder and ALS-based recommendation model which showed an improvement in impressions seen and duration spent by cold and sparse users.
TCS Innovation Labs - Research Engineer
◦ Developed attention mechanism architecture for detection of humor in edited news headlines using BiLSTM, knowledge graph and other extracted features. Paper published at COLING - 2020 conference in SEMEVAL workshop. Paper can be found here.
Quadratic Insights Pvt. Ltd. - Data Scientist
◦ Developed a hierarchical algorithm using text mining techniques and Naive Bayes algorithm to automatically redirect the customer complaint emails of a leading bank to the respective departments at three hierarchical levels.
Hindustan Petroleum Pvt. Ltd. - Operations Officer
◦ Validated machine learning models and developed new features to forecast sales of different oil products that helped in running the distribution plan effectively.