Experiences

• Built an AI agent from scratch using Amazon Bedrock to extract insights from EV&O (Causal Learning) model outputs, enabling economists, Product Managers, and scientists to analyze large-scale Excel data efficiently.

• Designed modular workflows integrating LLMs and programmatic reasoning. Ran thorough evaluations to ensure accuracy, robustness, and stakeholder alignment

•  Pioneered the integration of Generative AI and Cognitive AI to develop a robust Self-Explanation mechanism in VERA, enhancing interactive learning in ecological modeling for adult learners.

• Architected and integrated Theory of Mind Knowledge Models to develop a unified metacognition framework, consolidating disparate cognitive theories to enhance coherence across educational technologies.

•  Evaluated LLMs for the hard-constrained task of Program synthesis by Example (PBE)

• Designed novel techniques to increase the reliability and robustness of LLM-generated code against prompt sensitivity

• Developed a pipeline leveraging symbolic methods to repair inaccuracies in LLM-generated code

• Successfully solved over 30% of previously unsolved benchmarking programming tasks, demonstrating effectiveness across various programming languages, including Python and Excel

• Khan, R., Gulwani, S., Le, V., Radhakrishna, A., Tiwari, A., Verbruggen, G. LLM-Guided Compositional Program Synthesis. arXiv:2503.15540, 2025. 

• Revamped commenting functionality for a web-based review and approval workflow of structured technical documents

• Ideated and developed an extension framework for client customization without legacy code knowledge

• Prepared Mudra, a systematically curated multimodal Hindi co-speech gesture dataset of 43.62 hours of videos and extracted poses, transcripts and audio features

• Generated culture and gender influenced co‑speech gestures for a conversing virtual agent

• Conducted pilot studies and statistical analysis to confirm gender and culture influences in co-speech gestures

• Used CNN & Deep Feedforward neural network to extract and represent gender as a spectrum

• Achieved 42% improvement in the accuracy of existing AI used for accessibilty tagging of PPTs using human-in-the-loop techniques

• Improved the model accuracy by 30% for brochures by building a novel pipeline to propagate the human feedback and improve the model performance iteratively

• Built a prototype of the pipeline using PyTorch and Flask

• Estimated the uncertainty when answering queries in the absence of a true causal graph

• Compared and analysed the various structure learning algorithms and evaluation metrics and built a query based ensemble to output best algorithm given a query on a dataset

• Demonstrated a speedup of 88.65% over the existing strategies for runtime verification of JavaScript programs

• Designed and implemented specification and instrumentation mechanism to allow dynamic enforcement of properties and to ensure correct behavior and detect deviations

• Interviewed journalists at Press Club of India, Women Press Club of India and other news-rooms to gather insights about Digital Journalism in the 2019 Elections in India

• Developed a new understanding of the key drivers for social networking and digital media practices of journalists in India