Thathsarani Sandarekhathathsaranisandare

Partnered with startups and AI-native teams to design and ship production LLM systems across multi-agent orchestration, document intelligence, fine-tuning, and LLM safety. Built production-grade agentic routing between NL→SQL and hybrid RAG (BM25 + vector + reranked) using LangGraph with RAGAS-based evaluation and SQL safety guardrails. Engineered a hierarchical multi-agent legal AI system with self-refinement loops and automated citation verification achieving a 0% fabrication rate in testing. Fine-tuned Llama 3.1 8B using LoRA and DPO, then optimized inference with AWQ 4-bit quantization and vLLM batch serving. Delivered defense-in-depth LLM safety for regulated FinTech use cases using prompt injection detection, PII redaction, schema enforcement, hallucination detection, and auditable evaluation across 500+ adversarial prompts via Langfuse tracing.

Contributed to analytics and AI initiatives across hospitality and enterprise business units. At Cinnamon Hotels: helped build personalized offers & pricing systems, a Customer Lifetime Value prediction engine, and demand-elasticity-based pricing optimization for F&B and room products. Analyzed hotel energy usage by correlating occupancy, weather, and F&B revenue to identify optimization opportunities across 75% of total consumption. Supported food & beverage cost optimization (requisition standardization and interactive cost simulation dashboard) and integrated LLM-assisted reasoning into Azure Databricks pipelines using Azure OpenAI to reduce report generation time. Built 20+ PySpark/Databricks pipelines for dashboards and ML use cases, supported data quality engineering, statistical testing, and model development.