Nikhil Sharma

Data Scientist [Agentic AI at Pepsico] at Grid Dynamics (2025-04 – Present)

Architected and developed a multi-agent Agentic RAG pipeline (Temporal, Azure AI Search, specialized agents) with LLM-driven query decomposition and iterative self-reflection, improving answer coverage by ~35% while maintaining <30s E2E latency across 2–5 parallel sub-question searches per query.
Built a multi-agent orchestration platform on AWS Bedrock AgentCore enabling dynamic deployment of CrewAI and LangGraph agents via a unified FastAPI REST API, async two-phase CI/CD (Docker, ECR, CodeBuild), and AstraDB persistence.
Integrated MCP Gateway and Bedrock Memory to deliver context-enriched conversations with semantic search, short/long-term memory retrieval, and dynamic tool discovery — backed by OpenTelemetry observability (AWS ADOT), multi-tenant access control, and Pydantic v2 validation for production reliability.
Built a multi-agent orchestration platform using FastAPI, CrewAI, and the A2A protocol, routing ITSM queries to specialized ServiceNow AI agents for incident summarization, change planning, risk assessment, and trend analysis — with dynamic agent discovery, Redis-backed session management, and persistent multi-turn memory via Bedrock AgentCore across asynchronous workflows.
Architected a three-tier LLM/RAG evaluation pipeline using RAGAS, LLM-as-a-Judge, and Langfuse observability, that reduced hallucination rate by ~20%.

Data Scientist at TechStax Private Limited (2024-02 – 2025-03)

Built and deployed predictive models for Solder Paste Inspection and Pick-and-Place machines, reducing PCB manufacturing failures and minimizing raw material loss by 22%.
Developed an advanced LLM-powered chatbot using Meta's Llama 3, incorporating intent recognition, entity extraction, query construction, and conversational answer generation, enabling seamless natural language query resolution from databases.

Software Development Engineer (Machine Learning) at Capillary Technologies India Limited (2021-06 – 2024-02)

Enhanced and fine-tuned predictive models used in the back-end of a retail analytics dashboard, resulting in a 4% reduction in campaign costs for 45+ clients.
Developed a robust framework for evaluating propensity models aligned with specific business requirements, enabling accurate assessment of model performance.
Optimized the RFM model in production, slashing average model training time by 11% and generating significant savings on training resources.
Extracted key performance indicators (KPIs) from multi-dimensional datasets tailored to specific brands, leading to enhanced customer retention strategies.
Developed a tool to measure customer loyalty and assess the impact of loyalty programs on retention and sales growth, empowering customer success managers with actionable insights and user-level campaign personalization.
Developed and deployed a production application using OpenAI services to review and provide recommendations on pull requests raised for repositories.
Partnered with product managers and customer success managers to analyze KPIs across business verticals and build measurement criteria for customer loyalty and loyalty-program impact.

About