Dwij Patel

Designed and built a production-grade Voice AI agent using OpenAI, Twilio, Deepgram, and ElevenLabs. Designed conversational architecture with turn detection, interruption handling, and multi-step stateful dialogue management. Implemented queue-based state synchronization and parallel data extraction pipelines. Integrated structured LLM tool-calling to execute deterministic operations (scheduling, availability, calculations) beyond generative capabilities. Engineered IVR fallback system to handle failure cases in LLM-driven navigation, improving system robustness in edge scenarios. Developed evaluation framework to benchmark latency, response accuracy, and conversation success rate under production conditions. Built real-time monitoring dashboard with WebSocket streaming.

Selected for Antler US Residency Program. Built an LLM-driven analytics platform for brands selling on Amazon, Instamart, Zepto, and Blinkit. Designed scalable scraping infrastructure for Instamart, Zepto, and Blinkit. Developed embedding-based product matching system to normalize and map heterogeneous product data into a unified schema. Built distributed data pipelines using Celery for reliable ingestion and processing. Implemented agentic AI workflows combining embeddings, multimodal LLMs, and tool-based execution over ClickHouse.

Contributed to a production SaaS platform, building responsive frontend components and improving state management across complex UI flows. Led migration of legacy APIs to a modern Node.js (Express) architecture, including database transition to PostgreSQL with custom Python-based data migration pipelines. Optimized backend systems by transitioning synchronous workflows to asynchronous, event-driven architectures, significantly improving efficiency and reducing cloud costs. Worked across full-stack layers using modern web technologies, collaborating in both remote and hybrid environments.