Sep 2024 – Jul 2025 · 10 mos · Dubai, United Arab Emirates · On-site

• ▪ Designed and led the development of ZClip, an adaptive gradient clipping algorithm that improves stability and eliminates manual batch skipping during LLM training.
• ▪ Worked on initialization and variance control techniques to enhance LLM performance.
• ▪ Explored parameter-efficient Transformer designs using factorized embeddings and low-rank attention.
• ▪ Contributed to distributed LLM pre-training recipe design and built scalable training pipelines using FSDP, PyTorch Lightning, and DeepSpeed, applied in training 7B-parameter models.

Jan 2022 – Sep 2024 · 2 yrs 8 mos

• ▪ Co-Author of Komodo LLM: A foundational large language model tailored for a specific language.
• ▪ Trained and deployed a variety of task-specific Language Models (LLMs) in production, significantly improving the system's ability to comprehend user messages. 🚀
• ▪ Engineered a series of embedding and generative zero-shot models with distributed pre-training, enabling rapid bot creation and reducing unidentified utterances by 30%. 🤖✨

Dec 2020 – Dec 2021 · 1 yr

• ▪ Developed and fine-tuned custom GPT and BERT models from scratch using in-house chat data for content generation and embeddings.
• ▪ Developed a custom probabilistic spelling correction model leveraging our in-house corpus, enhancing accuracy in fixing domain-specific words.
• ▪ Developed a Conversational Sentiment Model utilizing deep learning techniques.

Jul 2018 – Dec 2020 · 2 yrs 5 mos · Bengaluru, Karnataka, India

• ▪ Working on a neural ranking model using deep learning with semi-supervised techniques.
• ▪ Built a Pairwise Job recommendation engine that recommends jobs to the candidate looking at the candidate's profile.
• ▪ Implemented Transformer AE, GPT-2, and a custom autoregressive transformer model from scratch in TF2.0, with distributed pre-training on a large corpus.

Sep 2017 – Jun 2018 · 9 mos · Bengaluru, Karnataka, India

• ▪ Did POC on context-aware information retrieval using semi-supervised learning for longer query text.
• ▪ Built multiple autoencoders for short and long sequences text documents using CNN and LSTM.

May 2016 – Sep 2017 · 1 yr 4 mos · Gurugram, Haryana, India

• ▪ Built an opinion Mining module using Deep Learning.
• Used a sequential classification model using LSTM to extract opinions from documents.
• Used Word2vec and Character-CNN for feature generation.
• Used Text CNN for relation extraction.
• ▪ Developed a NER module to extract entities from text documents.
• Built a sequential classification model to identify the entity among domain-specific documents.
• Used Word2Vec and K-Means Clustering for feature generation.
• Implemented Big-Data pipeline using PySpark for computation of Big-Data on Cluster of Computer.
• Used Natural Language Processing in conjunction with the acquired knowledge domain to extract features from documents.
• ▪ Developed an extraction module to extract phrases from text documents.
• Built a sequential classification model to extract phrase-based entity among the domain-specific
• documents.
• Used Word2Vec and K-Means Clustering for feature generation.
• Algorithms used: Neural Network, LSTM, CNN, Word2Vec, Conditional Random Fields
• Technologies or Technical tools used: TensorFlow, Core Java, Python, HBase, Apache Spark

May 2015 – May 2016 · 1 yr · Noida, Uttar Pradesh, India

• ▪ Built a sequential classification model to remove the junk part of the data.
• ▪ Used Producer-Consumer Design pattern in crawler to handle multi-threading.
• ▪ Developed a web crawler for crawling the legal documents from across the websites and used IP Switcher API.

Jan 2013 – Sep 2014 · 1 yr 8 mos · Noida, Uttar Pradesh, India

• Bindaaslo was a fully controlled online marketplace, promoted by Anyaani Technology Pvt. Ltd.