Hands-on engineer with experience in end-to-end development of machine learning systems.
I bring 10 years of diverse industry experience at a Document AI startup, Conversational AI startup and Display Advertising Team at Walmart.
Document Search - Developed a fast, accurate, and cost efficient document search engine using the multiple techniques like context preserving chunking strategies, generating document outline of headings and sub-headings for information location etc.
•
Document Agent Builder - Developed a ReAct prompting-based agent builder framework by which users can create use-case specific agents with the help of Weav tools library. An agent can perform complex and multi-step tasks involving searching documents, querying tables, searching the internet etc
•
Insurance Underwriting Copilot - Designed and developed the AI backend for underwriting assistant copilot. This copilot extracts data from submission documents and external databases and consolidates it to create a knowledge graph. Then, performs exposure analysis and loss runs review using Knowledge graph agent.
◦ ML based Agent performance evaluation: Led a team of three ML engineers. Project involved building text classifiers to identify specific agent behaviors (apology, empathy, confirmation, understanding-issue) and establishing their value in agent performance evaluation. Also, played a critical role in project planning and product design.
◦ Agent transfer detection: Ideated and defined the project to identify if agent transferred call to another agent. Built data labeling guidelines, developed model and integrated with production system as a REST API service.
◦ Lego v2 service: Developed a service that combines outputs of multiple models based on the product logic eg: combining money-entity detector model and loan-keyword detector model to confirm if agent stated the amount customer owed. It enables defining custom combination logic of multiple models and seamless onboarding of ML models.
◦ ML infra cost monitoring and optimization: Developed a tool to track ML infra cost that provides breakdown across various ML services and environments. Implemented two key measures - concurrent request processing using cooperative multi-tasking framework and efficient service scaling policies - leading to 43 percent reduction in ML service infra cost.
◦ Bidding strategies for display ad campaigns: Built ML models to predict site-revisit probability, purchase probability, and profit estimate for Walmart users based on site activity logs. Model prediction scores were used to decide bid to show ads to users.
◦ Ads impact measurement: Designed bias-free A/B testing strategy to measure the incremental purchases driven by display ads.
◦ ML interpretability: Surveyed state-of-the-art approaches of ML interpretability.
