Generative AI Developer
Reliance Jio Infocom Ltd.
Telecom RAG Knowledge Base Agent • Architected an end-to-end RAG pipeline over 5,000+ telecom documents using LangChain, ChromaDB (vector database & embeddings), and GPT-4, cutting engineer query resolution time by 60%. • Enabled multimodal understanding by transforming documents, diagrams, and architecture flows into semantic embeddings to improve retrieval accuracy and reduce hallucinations through structured data preprocessing and validation of structured and unstructured datasets. • Exposed production-grade FastAPI services with MCP integration for knowledge ingestion, retrieval, and agent orchestration; supported model deployment via CI/CD pipelines with monitoring. • Tech Stack: Python, FastAPI, LangChain, ChromaDB, Azure OpenAI (GPT-4), Nomic Embed, MCP Multi-Agent GenAI Platform – A2A Communication • Designed an Agent-to-Agent (A2A) communication framework enabling specialized AI agents to collaborate on complex telecom workflows across Jio's platform. • Built MCP-enabled microservices using FastMCP allowing agents to expose and consume tools dynamically, improving inter-agent orchestration and reducing integration overhead. • Engineered a real-time Kafka-based fault summarization engine generating automated LLM incident reports for NOC teams, reducing manual triage effort by ~70%; collaborated cross-functionally with data scientists and software engineers for integration. • Tech Stack: Python, FastAPI, FastMCP, LangChain, Apache Kafka, OpenAI APIs, MCP, A2A GenAI Network Configuration Assistant • Designed a conversational LLM-based assistant for engineers to query, validate, and generate device config scripts via natural language — cutting deployment time from 4 hrs to 25 min (89% reduction). • Indexed 10,000+ pages of vendor docs (Cisco, Nokia, Ericsson) into Pinecone via LlamaIndex for precise, context-aware RAG-based retrieval, with iterative prompt engineering and quality evaluation against defined metrics. • Tech Stack: Python, FastAPI, LlamaIndex, Pinecone, OpenAI APIs, LangChain Multi-Turn Customer Support Chatbot • Built a multi-turn GenAI chatbot (Google Gemini API + Django REST) processing 10,000+ queries/day with 85% deflection rate and prompt templates optimized for 15+ telecom use cases. • Integrated Hugging Face models (Mistral, LLaMA-2) as fallback LLMs, applying fine-tuning and model evaluation to cut API costs by 40%; deployed 6 microservices on AWS Bedrock & Azure OpenAI achieving 99.5% uptime with MLOps and model monitoring. • Tech Stack: Python, Django REST, Google Gemini API, Hugging Face (Mistral, LLaMA-2), AWS Bedrock, Docker
