Hi, I'm Sak Nagpal

Built and deployed a production LLM-based customer complaint classification and routing system processing 100k+ historical enquiries at Jetstar Airways. Combined the Claude API with KNN clustering to identify complaint categories, achieving 92% classification coverage — replacing a brittle rule-based approach and materially improving routing accuracy and resolution efficiency. Designed a RAG-based architecture integrating internal knowledge sources, evaluating trade-offs across latency, cost, and model performance.

Designed and delivered an NLP pipeline for automated classification of aircraft maintenance and engineering logs at Jetstar Airways, where unclear technician-written entries previously required manual expert review. Extended the system to forecast recurring fault patterns across the fleet — reducing mean time to identify faults by 30% and enabling proactive maintenance scheduling. The system handles noisy, domain-specific language with high accuracy across a diverse range of fault types.

Architected and built an internal LLM-powered knowledge assistant using Amazon Bedrock, trained on operational manuals and staff travel policy documents. The system enables accurate, instant retrieval of policy information — reducing staff onboarding time by over 95% with near-100% accuracy in seat allocation decisions. Built on a RAG architecture with careful chunking and retrieval strategies optimised for long-form policy documents. Currently in final pre-production validation.

A fully agentic React + FastAPI application that autonomously researches Melbourne suburbs for property investment decisions. Built on a LangGraph state graph with conditional edges and real-time streaming SSE — the agent selects tools dynamically with visible reasoning steps. Powered by live data sources: property prices and median sale stats from Victorian Valuer General (VGV) sales data, and suburb demographics from the ABS 2021 Census DataPack covering population, median income, household composition, and owner vs. renter split. Price trend queries render an inline Recharts chart directly in the chat UI. A Haiku-based query guardrail classifies and short-circuits irrelevant queries before they reach the agent loop. Amenities integration in progress. Backend deployed on Railway, frontend on Vercel.