| CARVIEW |
Biography
Hi, This is Nikhil Madaan. I’m a Senior AI Research Engineer at Bloomberg AI, focusing on multi-variate time-series modeling to build real-time pricing models.
Previously, I was a M.S. in Computer Engineering (AI-ML Concentration) student at Carnegie Mellon University. During my time at CMU, I was a part of MultiComp Lab, where I worked with Jianing “Jed” Yang and Prof. Louis-Philippe Morency on bias analysis in Multimodal QA datasets. I was also a part of Lion’s Research group, where I worked on Personalized Federated Learning with Dr. Taejin Kim and Prof. Carlee Joe-Wong.
During my master’s, I interned at Amazon as Applied Scientist, with the Media and ADs group, where I worked with Dr. Manisha Verma on multi-modal product headline generation. Prior to my master’s I worked as SWE-ML at Flipkart as a part of the Catalog Ingestion team, where I played a significant role in designing and implementing scalable AI systems.
I am particularly interested in the applications of Deep Learning in areas such as Natural Language Processing, 3d-Computer Vision, and Multimodal ML.
- Multimodal ML
- NLP
- 3D Computer Vision
- Robotics
M.S. in Computer Eng. (AI/ ML Concentration), 2021-2022
Carnegie Mellon University
B.E. Hons in Electrical and Electronics Eng., 2015-2019
Birla Institute of Technology & Science, Pilani
Experience
- Working on leveraging ML models for multi-variate time-series modeling to build real-time pricing models for Fixed-Income securities.
- Worked on generating headlines for products by factoring in multiple modalities such as Product Images and product attributes (Text); using SOTA multimodal fusion networks such as Flava, Mantis.
- Employed contrastive learning to improve the diversity of the generated headlines and rouge, bleu score by 53.5% and 145% respectively, w.r.t unimodal models.
- Worked on analyzing QA bias in Multi-modal Question Answering Systems using fine-tuned language models.
- Worked on the transferability of adversarial attacks in Personalized Federated Learning setup and trying to make the learnt model more robust to such attacks..
- Leveraged Image encoders such as ViT, ResNets, to generate embeddings of the product images. Indexed the generated embeddings, added multi-cluster support for reads and writes, and used the embeddings to support Product Deduplication.
- Developed a prioritized distributed message processing xtension to the camel-Kafka component in Java, to support priority consumption of records and implemented various consumption strategies to support different use cases.
Recent Publications
Accepted at ICRA 2024 | LangRob @ CoRL(LangRob) 2023.
Accepted at KDD (Multimodal) 2023.
Accepted at Neurips 2023.
Accepted at SLT 2022
