Nikhil Soni

About

Built and deployed a production LLM-powered market intelligence and decision-support platform for noisy, real-time financial data, using a high-throughput embedding + vector DB RAG architecture to reduce hallucinations while keeping latency and cost low. Operated it at scale with GPU-backed inference (continuous batching/quantization), FastAPI on Kubernetes, and Airflow-orchestrated ingestion/embedding/retraining workflows, with strong schema-based reliability and monitoring.

Key Strengths

• Built and deployed production LLM-based market intelligence/decision-support system
• Designs RAG pipelines to reduce hallucinations and ground outputs in real-time data
• Optimizes LLM serving for latency/cost at scale (GPU inference, continuous batching, quantization)
• Improves reliability via structured outputs, JSON schema validation, retries, and graceful degradation
• Strong observability mindset (monitors latency, token usage, and response-quality signals)
• Uses offline test suites and replay of real inputs to evaluate correctness, latency, and cost
• Runs controlled experiments when changing models/prompts to manage accuracy vs. cost tradeoffs
• Effective cross-functional collaboration translating non-technical goals into system requirements

Experience

Education

Awards

• Merit-based scholarship - New York University

Skills

Languages