sridhar.d.kamoji

• Architected and deployed an end-to-end Agentic NL2JSON system converting natural language requests into complex deployment-ready JSON workflows; the system autonomously fetches resources, generates pseudocode, and executes multi-step reasoning and self-healing to generate the JSON workflows, achieving 89% avg workflow generation accuracy and 95% intent coverage; built on LangChain, LangGraph, PostgreSQL, Django, Docker, AWS EKS. • Engineered a robust automated test evaluation framework to validate JSON workflow accuracy and reliability at scale, ensuring production-grade quality.

• Designed a multi-agent Plan-and-Execute LLM Compiler that decomposes complex financial queries into parallelizable sub-plans, routes them to specialized downstream agents, and synthesizes comprehensive responses; built on GPT-4o, LangGraph, FastAPI, Docker.

• Built a Customer Support AI Chatbot (RAG + agentic) integrating Llama and Qwen2.5 models; achieved 93% top-7 retrieval precision and 1.7s average latency via domain adaptation (BERT/SBERT), OOD detection, and custom DataDog monitoring. • Developed an internal agentic chatbot over Ericsson GAIA’s Jira/Confluence/GitLab data using tool-calling LLMs and LangGraph-based agentic RAG pipelines. • Built and exposed a production PDF parsing microservice (header/footer removal, TOC & section extraction) via FastAPI, Docker, and Kubernetes for the ELI RAG platform.

• Developed a customer propensity model on 480M+ records (XGBoost) to predict business bank account opt-in; achieved 85% AUC, 2.3x lift, driving a 19.3% conversion rate increase and 27% reduction in marketing costs. • Led large-scale data migration (˜40TB) with PI masking using PySpark for model development for external company - Quantium.

• Deployed anomaly detection models (PCA reconstruction error, Mahalanobis distance) for semiconductor manufacturing tools, identifying cryogenic device failures and delivering more than $100K/year in savings. • Built multi-class text classification system for manufacturing non-conformity assignment, reducing manual effort by 30%; built TFIDF/SBERT-based information retrieval system for engineering support.

• Delivered high-value customer propensity models (Random Forest) for Beauty Systems Group LLC and Lamps Plus; achieved 91–93% AUC and 2–2.5x lift in identifying future high-value customers. • Developed a 90-day churn prediction model (Logistic Regression) for Beauty Systems Group LLC; achieved 87% AUC and 2.3x lift, enabling proactive retention campaigns.