Jayant Duneja

- Knowledge FAQ Generation: Built a recurring pipeline to generate high-quality FAQs and knowledge base recommendations from daily call transcripts. The system extracts and filters question–answer pairs, clusters them incrementally against existing knowledge, and identifies inconsistencies across generated answers. Implemented hallucination filtering and controlled execution to ensure reliable outputs at scale. - Scaling & Reliability: Improved service throughput and stability by fixing backpressure handling in a Reactive Java service and propagating it correctly to the entry point. Enabled autoscaling based on internal queue depth to regulate load under traffic spikes. Increased capacity from ~300 requests per minute to ~6,000 concurrent requests while maintaining system stability. - Talk to Your Data: Built a conversational analytics prototype that allows users to query call analytics and transcripts using natural language. Designed a multi-agent LLM architecture with orchestration, semantic querying via Looker, and scalable transcript analysis. Implemented session management, context window reconstruction, and real-time streaming responses.

- Classification Improvement: Redesigned the call classification pipeline to significantly improve accuracy from ~40% to ~80% by extracting richer signals from call transcripts. The system now uses both topic and one-line reason-for-call summaries, replacing short topic-only labels that produced generic clusters. Introduced embedding-based cluster shortlisting to reduce prompt overload and improve LLM classification reliability. - Custom LLM Prompt Integration: Developed a backend solution to empower customers with the ability to define, test, and deploy custom LLM prompts for transcript evaluation, seamlessly integrating results with both the data warehouse (BigQuery) and the primary database (Google Datastore) to deliver real-time, data-driven insights in Looker to 50+ dashboards.

- Designed and implemented a scalable event-driven architecture for data pipelines using Java, Spring Boot, Google Pub/Sub, and Azure Data Factory REST APIs. Automated approximately 75 ETL workflows by subscribing the service to multiple Pub/Sub topics, resulting in a 10% increase in client efficiency. - Improved operational robustness with a secure API key management feature using HashiCorp Vault, enabling teams to efficiently handle API secret keys for diverse development environments and enhancing security across the system. - Implemented 95% code coverage by adding integration and unit tests utilizing Sonarqube and Junit. Responsible for maintaining the entire operational load involving new features, performing software upgrades, and fixing bugs. - Successfully navigated user resistance to new messaging formats by onboarding new clients through inter-team communication and providing documentation.

- Developed and tested a CLI tool in Python with the help of Splunk SDK to reduce lookup time and automate the examination of error logs on Splunk Dashboard utilizing API Calls. - Implemented granular filtering of error logs by platform, timeframe, and error ID, which simplified error analysis for developers with diverse technical backgrounds. - Deployed the CLI tool employing Flask on Walmart Cloud leveraging CI/CD tools, utilizing Docker images and Kubernetes deployments for scalability and ease of maintenance.

Relevant Courses : Object Oriented Analysis and Design Datacenter Scale Computing Natural Language Processing Neural Networks and Deep Learning