Jul 2023 – Jun 2025 · 1 yr 11 mos · Seattle, Washington, United States · Remote

• Building the DeepScribe Agent
• Scoping and building provider facing features like auto-generation of After-Visit Summaries.
• Led the team building workflow orchestration tools enabling higher engineering velocity scalable ways to build multiple medical specialty models.
• Led the development of an Auto-Evaluation framework consisting of statistical and LLM based evaluation methods that helps domain experts evaluate model output quality. This improves trust in model pipelines and shipping velocity.
• Led the team building next-gen medical AI solutions for Value Based Care - leading to the launch of HCC Assist [https://www.deepscribe.ai/hcc]
• Led the Instruction Fine-Tuning team to design the next-gen proprietary Medical LLM focusing on EHR-compatible automatic note creation.
• The achievements of the team included the deployment of our in-house GPT4-independent LLM and pipeline that creates notes with better quality than GPT4

Aug 2022 – May 2023 · 9 mos · Seattle, Washington, United States

• Modeling: Designed and trained SoTA DL - MMOE and PLE based models for personalization of lenses on different surfaces - also creating and choosing language and visual features.
• Models showed improvements of user engagement ranging from 2% - 100% for different surfaces - translating into xx% increase in revenue.
• Data Infrastructure: Built training data creation pipelines which process 1 peta-byte of data. Further built ETL pipelines to create ingestible TF features.
• Uses Apache Airflow utilising Apache Beam (through Google Dataflow) and BigQuery to merge features (logged from 100K qps to the service) and impressions.
• Conducted AB tests to optimize sampling - reducing data processing costs by 33% while also increasing engagement by about 1-3% (~5M users)
• Model and Feature Monitoring / Observability: Conceptualised and implemented a framework to identify, measure and alert for any anomalies in model prediction (behavior) and feature distribution.
• Used by the entire team - Helps capture broken feature pipelines and stale models to avoid technical debt and regressions.
• Improving the freshness of lenses: Designed internal diversity metrics to measure freshness of lenses using visual lens embeddings and tag embeddings.
• Used by the entire team - to measure the improvements (through AB tests) in the user experience and engagement on the app.

Jun 2021 – Jul 2022 · 1 yr 1 mo

• Extended the applications of CASPR to cohort-identification, fraud detection and out of the box feature-creation.
• Conceptualized and Developed a News Classification and Ranking service within Supply Chain Insights that helps our customers (businesses) to be prescient about potential risks to their supply chain based on world news.
• Deployed the model to a real time endpoint in Azure within a span of 2 months - after performing extensive load testing and adding optimisations - reduced the 429 errors to zero.
• Worked with the Product Recommendation team to develop DL based collaborative filtering models to be used by various consumer faced brands. Also leveraged CASPR to work with techniques like ALS and Random Forest to develop hierarchical models.
• Worked with LightGBM on Spark to train glassbox models on 100 million rows of data generated predictions.

Nov 2019 – May 2021 · 1 yr 6 mos

• Designed a patented Deep Learning technology called CASPR that can generate user representations based on their activity (sequential) data. The algorithm works best for data stored in relational databases.
• Adapted CASPR to work with large amounts of data by using Spark for distributed ETL and using Petastorm and Horovod for distributed training on GPU's.
• CASPR beats other SoTA architectures for representation learning on structured data - we also fine-tuned it for use cases like: Churn Prediction, CLV prediction and Product Recommendation.
• Led the responsible AI charter in the team for DL models by creating an explainability library to explain Deep Learning models (starting with CASPR) leveraging modern research including Shap, DeepLIFT, LIME and others - by utilizing Captum.
• Researched and developed GAM-like LGBM-based architectures to create glassbox explainable (fully transparent) ML models which are usable in sensitive industries.

Jun 2019 – Sep 2019 · 3 mos · Germany

• Learning Subgraph Embeddings via a Subgraph Proximity Measure
• Published in an A-Level Data Mining conference : PAKDD'20 (Link: https://link.springer.com/chapter/10.1007/978-3-030-47426-3_38)
• Representation learning on large graph networks to identify subgraphs which are similar to other subgraphs
• Useful for tasks like Link Prediction, Node Clustering and Cascade Prediction
• Exploits the idea of Personalised Page-Rank to determine similarity of every pair of nodes and uses a novel weighted measure designed to eventually calculate the subgraph similarities.

May 2018 – Jul 2018 · 2 mos · Bengaluru Area, India

• Joint Intent Detection from complex sentences and slot filling, Bixby Team
• Was awarded the Pre-Placement Full Time Offer for commendable results.
• Designed a model that can be embedded in robots for them to be able to understand commands better by enhancing slot filling for each word in the sentence
• Prepared a RNN based Encoder-Decoder sequence model using a Bi-LSTM for Slot Filling and Intent tagging for input sentences.
• Added Attention layers in the Decoder to increase the accuracy.
• Modified the model to work for complex and dual intent sentences by first separating them and then feeding it to the model
• Generated 'iob' tags for each word in the sentence, by using ATIS data set.

May 2017 – Jul 2017 · 2 mos

• Semantic matching of text using Deep Learning models, Research Division
• Designed a model for semantic clustering of different kinds of questions to be embedded in chat bots for customer question-answering
• Developed a Deep Learning model using embeddings and Simple Neural Network layers
• Experimented with the addition of features such as Dropouts and Normalisations to prevent overfitting
• Used the Quora "Question-pair" dump and achieved an accuracy of 77%.

Jan 2017 – Apr 2017 · 3 mos · Assistech Labs, IIT Delhi

• Tactile Reader Application to facilitate studying for Visually Impaired
• This was done as an Embedded Design project, in IIT Delhi, mentored by Prof. M. Balakrishnan in 2017.
• We were a team of 6 people who designed a Desktop IDE in Visual Studio, and an android app.
• Improvised to develop a click detection mechanism through a gesture using real-time finger scanning.
• Designed a Visual Studio software for semi-automated generation of data to be used by the Android App which used DCEL structure
• Blog At: audiotactilereader.wordpress.com