Jun 2021 – Sep 2025 · 4 yrs 3 mos · Austin, Texas & Toronto, Canada · Hybrid

• Designed and deployed cutting-edge AI models in Oracle Cloud Applications.
• Representation learning for recommendation and ranking: Implemented pipelines
• with LLMs, word/document embeddings, Graph algorithms to generate robust
• representations of text entities. > 5% increase in r-precision in products like
• candidate to job matching in Oracle cloud applications.
• Domain specific NER pipelines for recommendation: Implemented Entity Extraction
• (NER) models using LLMs and deep learning architectures for extracting domain
• specific entities used for recommendations. Worked on: Data curation, supervised
• Fine-tuning, Evaluation framework for monitoring performance metrics
• (Precision@K, NDCG), robustness metrics (WEAT score), distribution shift metrics
• for 100+ customer datasets (100 mil+ records).
• LLM Evaluation Framework : Implementations for evaluating LLMs in production,
• supporting various metric types: LLM-as-a-judge: toxicity, bias, traditional NLP
• metrics (e.g. ROUGE, METEOR, BLEU) for various text-enhancing use-cases in
• production using LLMs (CommandR, ChatGPT, LLama).
• Neural Rewrite for Bias Mitigation: Implemented a text correction system to neutralize biased text in documents using BERT for bias detection and BART for neutralized text generation.

Mar 2021 – Aug 2025 · 4 yrs 5 mos

• Collaborate on independent AI Research related to computer vision and Natural Language Processing.
• Led to publications at ICLR, ICCV workshops and EMNLP conference.
• 1. Bias in Named Entity Recognition for Occupations through Large Language Models
• ICLR 2025 - Workshop on Spurious Correlation and Shortcut Learning: Foundations and Solutions
• https://openreview.net/forum?id=xfCjvr8MWR
• 2. Intersectional Fairness in Multi-Label Chest X-Ray Classification
• ICCV 2024 Computer Vision for Automated Medical Diagnosis (First Author - joint)
• https://arxiv.org/pdf/2403.18196
• Designed a cost-effective method to fine-tune multi-label computer vision models for fairness across race, income, and health insurance attributes by adapted subgroup robustness methods and loss-functions to handle class imbalance and multi-label fairness constraints.

Jun 2020 – Aug 2020 · 2 mos · Yorktown Heights, New York, United States

• AI Compute Division, IBM Research
• Worked on research and implementation of efficient AI architectures.
• Designed novel efficient neural network architectures
• using Boolean neurons optimized via mutual information-based feature selection.
• Built tree-structured Boolean models by recursively partitioning feature space (similar
• to decision trees) using Boolean Neural Networks.
• Applied multi-bit Hadamard encoding for multi-class tasks.
• Achieved 200× model size reduction with accuracy comparable to Random Forests; networks easily converted to compact hardware circuits (Sum-of-Products/LUT-based).

Sep 2019 – Mar 2022 · 2 yrs 6 mos · Buffalo-Niagara Falls Area

• Few shot multi-label learning on novel classes 200k (700GB) medical images. Proposed novel method combining multi-label prototypical networks, multilabel neighborhood component analysis loss, and fine-tuned models using a residual network backbone. The novel approach using a geometric ensemble technique resulted in 0.75 AUC on novel few-shot classes. MICCAI’22 ( DALI )
• Implemented new deep learning algorithms related to Federated Learning and knowledge distillation in neural networks for resource-constrained settings and multi-modal data. ( Long paper at DS’2021 )

Aug 2017 – Aug 2019 · 2 yrs · Greater Bengaluru Area

• ⚬ Implemented AI models at a large scale on 300 million users to predict risk for repayment of digital credit. This resulted in 1 million+ new users while incurring low losses at <5%.
• ⚬ Engineered Spark-based distributed pipelines to generate 1500+ features, to transform unstructured data sources: Text, Geolocation using NLP, and alternative data sources that capture predictive signals.Deployed Machine learning scores for production use through Elasticsearch REST APIs. Other Projects: Early Fraud detection, User targeting.

Jul 2016 – Aug 2017 · 1 yr 1 mo · Bengaluru, Karnataka, India

May 2016 – Aug 2016 · 3 mos · http://www.shogun-toolbox.org/

• Selected twice for Google Sponsored Summer Program. Mentor: Heiko Strathmann, PhD, UCL London
• Implemented efficient ML algorithms for an Open Source Project. Added new implementations using state-of-the-art research publications and approximate algorithms.
• ⚬ Improved existing algorithms through parallel processing using OpenMP, Vectorized implementations using Eigen3, and GPU-based computations. Algorithms: LARS, Boosting, KMeans++, Random Forest, Locality sensitive hashing for fast NN searches. Results were up to 5x compute efficient algorithms and 2x performance.