Mar 2023 – Jan 2025 · 1 yr 10 mos

• Led a team of 4 ML Engineers and 3 DevOps engineers; served as the critical integration point between DS, Platform Bidder, DevOps, and Data Engineering.
• Shipped 12 advertiser-specific micro-models with robust pipelines, achieving 10% CPI improvement with a framework scalable to 100+ models without manual intervention
• Drove 3x deployment throughput (9→28 models/year) through pipeline optimization and process improvements
• Developed session depth, device fingerprinting, and budget pacing features powering new data-centric optimization strategies

Mar 2021 – Feb 2023 · 1 yr 11 mos

• Optimised training/shipping pipelines such that the number of ML model deployments grew 3-fold from 9 models in 2021 to 28 models in 2022, when our scale was growing at a tremendous pace and DS:MLE ratio was still 5:1
• Re-organised & optimised numerous data pipelines and legacy codebase to fit the changing needs of the ML algorithms while adding along the way features like CPA-Retargeting optimisation models, enhancing A/B testing framework for flexible experimentation, progressing the CPI optimisation models, and bifurcating models by traffic type, segment etc
• Worked on deploying 4 major models : User Counter, Taxonomy & App Categories, Campaign BusinessType, & User level feature model which cumulatively gave a bump of 25% in desired CPI & IPM metrics
• Improved the resilience of the existing pipelines with enhanced monitoring, alerting and integrations which saved ETAs on Incidents detection and resolution 2x faster than back in early 2021.

Apr 2020 – Mar 2021 · 11 mos · Remote

• Mentored and taught over 250+ data science aspirants, under Machine Learning, Recommender Systems, Deep Learning & Natural Language processing course.
• Mentored a batch of 15 Industry Leaders who were in their career transformation journey into AI. Guided them in enabling and identifying AI/ML opportunities in their current role & also helped them with an overall industry outlook.

Feb 2020 – Feb 2021 · 1 yr · Ahmedabad, Gujarat, India

• Led the remote product roll-out in Alpha stage in 15 countries during onset of Covid in March 2020, across the largest Shipping Line in the world, leading NVOCC in India, and top AirForwarder in India. Took charge of Customer Onboarding, User Acceptance Testing, Deployment strategies, training & documentation material culminating in 40 daily active customers within 4 months.
• Designed a feedback loop mechanism for continuous training of ML models, which would capture the corrections made by the end-user and pass them as training data in a stratified manner. This enables the models to stay up to date and counter data drift in production
• Worked on cutting edge DocumentAI solutions to utilise the multimodal transformer-based architectures to solve problems like key-value pair extraction and mapping with advanced architectures like LayoutLM, LayoutLMv2, DocFormer, etc.

Feb 2019 – Feb 2020 · 1 yr · Ahmedabad, Gujarat, India

• Led the AI product roadmap, strategy, technical guidance for both Document Extraction product and the AI needs for Shipmnts Suite
• Built an end-to-end solution using custom YOLOv3 inspired architecture for object detection and localisation along with an orientation detection model followed by a classifier pipeline to process information from complex non-traditional images (stamps, logos, etc).
• Delivered an end to end product feature on document classification for limited data, this project included in-depth benchmarking of various ML and DL approaches for document classification and concluded with two scalable approaches, one is a custom CNN based architecture on the text and the other is a custom CNN based model which utilizes both texts as well as an image as inputs. This was deployed as a product where customers can seamlessly train models by just uploading the data.
• Designed and implemented an end-to-end platform for page stream segmentation (pagination) based on a custom deep network that utilizes both texts as well as image features to compare if two consecutive pages are a part of the same document or not. The entire system which has this as one of the components is now under Patent review status.

Feb 2018 – Jan 2019 · 11 mos · Ahmedabad, Gujarat, India

• Worked on Document Condition assessment from Image Perspective. Owned and delivered multiple Deep and Machine Learning models from inception to the Customer Feedback stage that involved semi-automated training of all Machine Learning models part of the solution
• Lead team for creating internal ML framework for automating rudimentary ML/DL tasks for Non-ML personnel in the company
• Lead Computer Vision team for our product to solve Document Processing problem statements from Computer Vision perspective. ie, Table extraction from documents - from Dataset collection to Experiment tracking using MLflow and DVC till deployable application using tfserve. I had worked on State of Art architectures like Yolo, EfficientDdet for detection and Segmentation use cases.

Jan 2017 – Jan 2018 · 1 yr · Ahmedabad, Gujarat, India

• Being the first Data Scientist on board, I explored the numerous areas where repetitive processes were rampant with Forwarders & CHAs be it email, physical documents, or phone calls for quotations
• Understood the pain points of Freight Forwarders, Clearing House Agents on the mundane yet important documentation work they have to do for creating Customs & Carrier Documentation
• Built an NER (Named entity recognition) pipeline for information extraction from Emails to speed up the process of Shipment Creation
• Performed EDA (Exploratory Data Analysis) on the Shipping documents to extract important insights using rule based techniques

Jun 2016 – Jan 2017 · 7 mos · Ahmedabad Area, India

• Worked on a project ContractSifter, a distributed web-application to extract information from legal documents, for our client LegalSifter.
• My responsibilities included:
• Pre-processing, cleansing, and verifying the integrity of data used for training.
• Text Mining using unsupervised and semi-supervised models.
• Optimizing classifiers (sifters) using machine learning and feature engineering.
• Building a platform to enhance data-collection, annotation, and sifter development process.
• Implementing multi-GPU training in TensorFlow in our project

Dec 2015 – May 2016 · 5 mos · Mumbai Area, India

• Used Markov models for random text generation and applied them to build a novel, sensible question generator.
• Built a back-propagation Neural Network for Upper and Lowercase alphabet recognizer, which gave percentage correctness of the hand-drawn letter
• Developed an algorithm to determine a set of litmus questions that determine a child’s intellectual level in the Concepts from previous, before entering the next grade thus providing a customized child-centric blended learning.
• Worked with POS tagger, Parse trees, Dependency trees using Stanford’s Core NLP package to build pragmatic question templates. Involved Answer word Classification, which was implemented through Decision Trees.
• Dived into Topic Modelling, Latent Semantic Analysis, understood the beautiful Math behind it, implemented it with and without popular packages like Gensim.

May 2015 – May 2016 · 1 yr · Chennai Area, India

• Implementing Bipolar Disorder of Basal Ganglia with Extended Reinforcement Learning Model
• Designed and implemented the classical Bandit and Grids World problem under varying concentrations of Dopamine & Seretonin
• Appreciated the above testbed scenarios in Value and Utility based Decision making
• Continued the work by modifying the update equations in pursuit of finding a faster medication than the already existing exercises

Mar 2014 – Aug 2014 · 5 mos · Chennai Area, India

• Implemented navigation of omni-directional y-chassis robot that gave about 30% increase in navigation
• Led the Recruitment Drive for Robocon 2015. Handled public relations, planned coding quizzes & conducted interviews to select 5 ace coders from 150 applicants.
• On-boarded new hires and mentored them on Robocon 2014's problem statement.
• Introduced Reinforcement learning as a training strategy and assisted in the initial design research and components procurement for Robocon 2015

Mar 2013 – Feb 2014 · 11 mos · Chennai Area, India

• Worked on Robocon 2013 contest's problem statement and ironed out bugs in the legacy code. (video linked)
• Learnt to use IMU, Rotary encoders and LiDARs.
• Optimized the hardware and software electronics of the 3 wheel pick and place robot with PS2 interface for manual driving.
• Contributed to the functioning of two robots according to the problem statement of Robocon'14.
• One was a manually controlled robot (Parent Robot) capable of picking and placing the autonomous robot (Child Robot) to specific places like see-saw, swing and pole walk. It was also capable of pushing one end of a see-saw & pushing the swing. It was driven on a 4WD Omni Wheel system and manipulated using pneumatic actuators. The Child Robot was capable of walking on a set of uniformly arranged poles using ultrasonic sensor based pole detection and pneumatically actuated grippers.
• Won the award for "Best Economical" Robot out of 92 teams that participated in India
• Won Best college club award at Aarush (SRM University tech fest) 2014's exhibition