Bringing the power of Large Language Models and Generative AI to the Enterprise.

􀁸 Tech Lead - Global Expansion of hyper-local Language Models (GeoLM):o DRI (Directly Responsible Individual) for developing location based speech recognition models (ASR) for the 10 largest iPhone markets servicing 3mn+ location related Siri queries dailyo Led direction, key design & pipeline implementation for 13 languages (morphological diversity), 10 countries (geographical diversity)o Featured in TechCrunch! Apple’s ML Spotlight! (Links below)􀁸 My team’s work on Knowledge Graphs for ASR was published at Interspeech 2021 (nominated for best student paper award): “A Discriminative Entity-Aware Language Model for Virtual Assistants”􀁸 Selected (1 of 24 from 10k+ Apple employees) to present graduate research at Apple ML Summit 2018,Title: “Representation Learning for Small Datasets” attended by 1000+ Apple employees (at the SteveJobs Theatre).􀁸 Tech Lead – Named Entity Analysis Toolo Developed a tool for systematically monitoring & analysing performance of ML models for highlycomplex low data languages (e.g. Italian, Korean, Russian, Turkish)o Helped surface the highest priority issues leading to >30% improvement in Named Entity ASRo Drove adoption across 10+ globally located language engineersTech: Python, Tensorflow, NLP, Distributed Computing (PySpark, Hadoop, Docker, Kubernetes), C++

Pragmatics􀁸 Developed a new ML algorithm for Representation Learning in small datasets & Identity Bias mitigation.o Benchmarked on Sentiment Analysis for Youtube videos with 8-12% increase in accuracy establishing new SoTA. Published in ICME ’17 (w/ Prof. Eric Xing & LP Morency): poster, 10 min. PPT, paper􀁸 Computationally measure Openness, Alignment & Topic development in Github discussionso Developed model to predict acceptance/rejection of pull requestso Achieved F1 score of 58.04, a 65% inc. over baselineo Key Results ⬇️ 5 min. readSemantics􀁸 StackOverflow answers search & retrievalo Compared GloVe & an early version of customized ranking embeddings (StarSpace framework)o Improvement of 60% on Hits@5 and DCG@5 metricso Integrated into the NLU component of a dialog agent assisting developer productivity􀁸 Trained Bidirectional LSTM on MS COCO dataset captions for caption generation conditioned on Inception V3 image features.􀁸 Compared Autoencoder, Variational Autoencoder (VAE), Conditional VAE & Deep Boltzmann Machines to model generative distribution of handwritten digits. (code)Syntax & Morphology􀁸 Trained bidirectional RNNs to predict POS tags on Brown Corpus, 96% accuracy, reproduced SoTA.􀁸 Implemented Unlexicalized Parsing & discriminative re-ranking to learn Probabilistic Context FreeGrammar Constituency trees over Penn Treebank corpus. (code)Special Projects􀁸 Trained a bidirectional GRU encoder/decoder model for symbolic/arithmetic reasoning of addition, subtraction. Achieved a MAE of 475 over 1300 test math problems.􀁸 Experimented with Bayesian Logistic Regression to build credible intervals for odds ratios of incomehypotheses on UCI Adult dataset.􀁸 Compared Deep Deterministic Policy Gradients & Normalized Advantage Functions to teach DeepMindMuJoCo’s ‘Half-Cheetah-v1’ agent navigation of a test obstacle course by training only for walking andjumping in simpler environments. (Report: Skill Trees for Hierarchical RL)

 Designed ML pipelines for identification of customer issues from call centre chats & social media formajor US telecom providers, healthcare chains (~2 million chats processed monthly) Developed a sentiment analysis toolkit adopted across the company to streamline/improve textualsentiment analysis. Geospatial Network optimization for improved logistics (R, GoogleMaps API)

Developed a unified in-house platform for mathematical and statistical analysis of large datasets (Java)

► Developed automation software for testing risk management applications. (Java)► The automation software was able to process and verify 11000+ test cases in 1 day covering all possible combinations of parameter values (compared to 800 test cases in 8 days previously).► Efficiency gain of 8 work days, increase in accuracy and robustness of testing

► Redesigned ‘Hedge Fund Management Game’, a business simulation game to train 60-100 students of the elective ‘Hedge Funds’ for the LBS MBA program ranked 1 globally. (Python, PostgreSQL, R and Django web framework)► Teaching Assistant in the course ‘Equity Investment Management’.

► Studied fundamentals of bonds, credit derivative swaps, credit derivative indices, specifically calibration of default probabilities, basic CDS pricing and credit curves.

► Formulated an investment thesis on a publicly-traded company and communicated the opinion through an equity research report.www.mentored-research.com

► Led a team of 40 people to organize the Out-of-Box Conference, a conclave of technology geeks, scientists, creative geniuses and business rebels. ► Based on the TED format► Instrumental in getting revolutionaries like Alphonse Kannanthanam, Vasu Dixit and Dr. P Namperumalsamy as speakers. ► The event was part of the international business plan competition, Conquest BITS, Pilani 2011.

► Led a team of 20 student athletes to participate in BOSM 2011 and won the gold cup for the best athletics team of the tournament.

► Prepared curriculum assignments and designed evaluation material► Assisted and supervised students in lab sessions for the course (With excellent recommendation)

► Analyzed technology and business trends in the web,social media, online services and software industry and prepared articles for WATBlog readers, followers and subscribers► Conducted interviews with business and industry professionals on behalf of the company for analysis and research

► Led a team of 2 programmers to develop a suite of software programs in C++ to simulate and optimize the Turtle Trading model by Richard Dennis, enabling statistical and heuristic analysis of large amounts of financial data.