Sumit Kushwah

Software Engineer

IndiaMART•  May 2022 - Present•  Noida, India

Developed a generative AI-driven RAG system leveraging the open-source LLaMA3 large language model to enhance the detection of banned products in seller catalogs on the IndiaMART platform, increasing precision from 60% to 90% and recall from 70% to 95% compared to the previous system. Orchestrated the deployment of Kubeflow and MLflow on Kubernetes, establishing one-click ML retraining pipelines for multiple in-house models, significantly enhancing experiment velocity with various datasets. Optimized Indiamart Buylead Recommendation model training code by leveraging GPUs with Facebook’s PyTorch, reducing training time from 5 hours to just 15 minutes. Deployed the AI4Bharat Indic Translation Model for translating product titles and descriptions from English to Indic languages, scaled to handle 100 requests per second with sub-100ms latency. Transitioned Indiamart Buylead auto-approval legacy system from PHP, Redis Queues, and VMs to a modern stack using Golang and Kafka Queues hosted on Kubernetes (Google Kubernetes Engine) with New Relic APM Monitoring, achieving a 97% reduction in memory footprint (from 315 MB to 9.5 MB). Engineered a custom parsing defense mechanism using Varnish Cache and Google reCAPTCHA to safeguard IndiaMART from parsing attacks, blocking 90% of bad traffic. Enhanced the Varnish Throttling module for IndiaMART by fixing bugs and adding new functionalities, contributing to open-source Varnish Cache modules. Compiled a report based on exploratory data analysis of buy leads marked as "not interested" from sellers, highlighting key caveats in the current buy lead recommendation system used by IndiaMART. Guided multiple teams at IndiaMART in leveraging AI coding tools like GitHub Copilot and Google AI Code Companion for tasks such as code generation, refactoring, and conversion, significantly enhancing productivity and code quality.

Software Developer Intern

Xorlabs.in•  April 2020 - October 2021•  Remote, India

Developed core front-end components for ’Vaizle,’ an in-house performance marketing tool, using Angular and integrating data from various social media marketing APIs. These components enabled clients to view insights and generate downloadable reports on their ad campaigns across multiple platforms. Developed a comprehensive Admin Panel for Fitelo, a health startup and Xorlabs client, using Angular for the frontend and Node.js for the backend. This system allowed Fitelo dietitians to efficiently manage and track weekly customer diets, serving as a central tool for diet management.