◦ AI Bench: Designed methodology & implemented runtime performance benchmarking with CI/CD infra on GCP.◦ Model Hub: Created LLM provider router supporting Perplexity, Replicate, OpenAI, Anyscale, Vertex AI etc.◦ Ivy Framework: Integrated Ivy into OSS repos - BentoML, Seldon-Core, Seldon-MLServer, MMPretrain. Added new functions, reviewed PRs, ensuring code quality & compliance with standards.◦ Transpiler: Remodeled source code generation logic to comply with torch.fx convention for easier debugging. Researched TPU support for transpilation & framework-agnostic (PyTorch, Jax, TensorFlow) profiler.◦ Enterprise Demos Lead: Collaborated with the Head of Operations on sales calls & market research for future product direction. Managed cross-functional teams, oversaw development, & defined demo objectives, ensuring alignment with organizational goals. Executed public demos showcasing value prop - XGBoost, MMPretrain.

◦ Inclusive AI: Addressed unintended bias in NLP, enhancing model robustness against adversarial text, surpassing Google & AWS sentiment analysis on internal datasets. Deployed inference API on AWS.◦ Health: Pilot project for predicting behavioral health from therapy sessions through multiple modalities.◦ Insights: Easy-to-digest information for therapists from user journals through Themes & Journal Abstraction AI.◦ ScreenerOMR: Constructed a Python package for extracting pertinent data from scanned medical screeners.◦ Annotation Framework: Implemented annotator cohorts of 10, & devised labeling process for golden data of sentiment, emotion, ACEs, and ICD-10. Wrote classification data augmentation with GPT-3.

◦ GenomicInfo: Engineered a framework for rapid genomic data extraction from C. Elegans papers for WormBase, employing a hybrid method of regular expressions, term dictionaries, & BERT NER.◦ Impact: Cut manual biocuration time from hours to minutes, while uncovering overlooked data from older papers. Published `Accelerated variant curation from scientific literature using biomedical text mining` in microPublication Biology with Valerio Arnaboldi, Paul Davis, Stavros Diamantakis, Magdalena Zarowiecki, & Kevin Howe.

◦ Masked Face Recognition: Crafted a facial recognition system compatible with masks for law enforcement authorities, tailored for low-end devices in collaboration with the image processing team.

◦ Detection: Developed custom object detection for low-end wearable devices in manufacturing - TFLite, Android.◦ e-Proctor: Built online exam e-proctor with facial recognition & features tracking, & gadget detection.◦ CADAnalyzer: Programmed engineering drawing comparison tool for universities.◦ Explorative POCs: Car damage diagnosis system on low data, & usage of privacy-preserving DL with PySyft.

◦ Analysis: Developed scripts to rectify discrepancies in geographical data for local dealers utilizing mapping APIs.