Sep 2023 – Apr 2026 · 2 yrs 7 mos · San Jose, California, United States · Hybrid

• Architected and deployed new document extraction pipelines using Graph Attention Transformers (vs. RF/MLP) for 10+ entities/20+ languages, increasing precision/recall by 5-10% and scaling Accounts Payable ARR by multiple millions by improving customer retention and deal win rates.
• ◦ Led organization-wide MLOps migration to AWS Sagemaker, reducing operational costs by 10x and eliminating custom GPU pod maintenance by standardizing the training pipeline.
• ◦ Designed and implemented fully automated Labeling and Training ML pipelines using
• Dask/Kubernetes/Sagemaker, enabling bi-weekly model releases, resulting in a 90% faster time-to-market for model improvements.
• ◦ Developed Private ZenLM, an adapter-based framework for AppZen’s Foundation Models, enabling the creation of custom, high-performance models per customer per entity enabling customer-specific fine-tuning.
• ◦ Built and deployed custom Computer Vision models (QR/Barcode/Payment details extraction) directly securing a major customer deal.
• ◦ Developing AI Agent Studio to allow customers to define custom Standard Operating Procedures (SOPs) in natural language via a chatbot, reducing the time required to define a complex
• SOP from 3 hours (manual) to under 10 minutes (an 18x efficiency gain), pioneering a new product line.

Mar 2023 – Sep 2023 · 6 mos · San Francisco Bay Area

• ◦ Built generic item detector using MaskRCNN and coupled it with custom few shot classification pipeline to correctly detect, identify and pick previously unseen objects with average rank 5.2 across 170 classes.
• ◦ Added partial checkpoint restore feature for all the training pipelines to support Tensorflow 2.x based models.
• ◦ Handled the transition of the instance segmentation pipeline to use Tensorflow 2.x and migrated the model and data storage from AWS based backend to GCP

Dec 2022 – Jan 2023 · 1 mo · San Francisco Bay Area

• • Working on Graph Neural Networks

Feb 2019 – Jul 2021 · 2 yrs 5 mos · San Francisco Bay Area

• Spearheaded projects to deliver cell nuclei dataset to the client Roche. Implemented PoC using CNNs, proprietary dataset generation algorithm and visualization tools. Generated ~250K in revenue.
• Automated and productionized tissue slide anonymizing tool to remove sensitive patient information on site. Achieved 15x speedup over the existing method.
• Proposed and implemented a CNN based multi resolution cell classification algorithm to detect different types of cancers.
• Built a cell segmentation algorithm by training Mask RCNN algorithm on limited dataset. Integrated this algorithm with the cell classification pipeline to create an end to end application for cell detection and classification.
• Created a highly scalable R-Tree based patching algorithm for batch processing whole slide images to reduce memory consumption by over 95%.
• Created an algorithm to cluster cells and construct concave hull to detect and compute area of tumor region. Computed infiltrating cells within the area to evaluate tumor status with 0.1um precision.
• Collaborated with senior pathologists to create a first pass tool using GANs to automatically ”score” certain slides. Increased scoring efficiency by 50%. (Patent Approved - US11158398B2)
• Developed a Named Entity Recognition tool using transformer networks to automatically parse research papers to identify certain elements. Reduced hours worth of manual effort to 1-2 mins.
• Implemented custom layers in Graph Neural Networks for predicting new edges in proprietary datasets.
• Designed ML interview process for the company. Hired and mentored ML engineers and interns for our cross-functional team across US and Asia.

Jul 2021 – Dec 2022 · 1 yr 5 mos · San Francisco, California

• Developed a custom text classification algorithm using transformer networks to auto-publish relevant news articles. Created deduping algorithm to post unique articles. Deployed at scale to serve thousands of users.
• Built conversational AIs to boost user engagement using custom GPT3 algorithms. Integrated it with Unreal Engine, Unity3D and private websites. Deployed on GCP to serve millions of users.
• Built an end to end application to create NeRF based 3D models from text prompt using Stable Diffusion and Instant NGP.
• Designed ML interview process for the company. Hired and mentored ML engineers and interns for our cross-functional team across US and Asia.

Aug 2018 – Dec 2018 · 4 mos

• Teaching Assistant and Grader for the graduate course CSE 512 Machine Learning co-managing a batch of 80 students.
• Tasks:
• Create new assignments, their solutions and assessing them.
• Create questions for the mid-term and end-term exams, their solutions and assessing them.
• Conduct office hours for discussion of students' doubts.
• Online solutions to students' doubts through Piazza.
• Topics:
• Probability, Regression, SVM, Regularization, Naive Bayes, Logistic Regression, Deep Learning, Convolutional Networks, Generative Adversarial Networks.

Jan 2018 – Dec 2018 · 11 mos

• Deep Learning, Computer Vision
• Built custom deep learning models for generating accurate density maps from crowd images using Generative Adversarial Networks in Python.
• Built a generator and discriminator architecture based on PIx2Pix paper. Experimented with the generator using custom U-Net architecture, Multi-Column CNN.
• Implemented Super Resolution as a processing step to increase model accuracy.
• Advisor: Dr. Minh Hoai Nguyen

May 2018 – Aug 2018 · 3 mos · Belmont, California

• Developed algorithms for preference prediction in dynamic environment with 99% test accuracy.
• Created proprietary Deep Learning based object detection models and deployed on resource-constrained devices while managing performance-accuracy trade-offs.
• Built a Deep Learning based NLP algorithm to detect trigger words from microphones.

Aug 2015 – May 2016 · 9 mos · Pune

• Implemented a generic binary classifier using suite of classification algorithms like SVM, Neural Networks, Decision Trees, etc. Used to classify patients suffering from brain disorders such as Schizophrenia and Dementia. Achieved AUC of 92% for Schizophrenia (119,748 cases, 2nd best in Kaggle) and 98% for Dementia (336 cases).
• Presented and published research paper in IEEE conference. (https://ieeexplore.ieee.org/document/7914961)
• Added support for multiple programming languages such as Caret(R) and Weka(Java) to allow users to efficiently use the models in their preferred language.

May 2015 – Jul 2015 · 2 mos · Pune Area, India

• Worked on JBoss Modularization of Flight Deck, a Maven project, that hosts 10 sub-applications (Spring, Java).
• Reduced the storage size of 10 deployed applications by around 70%.