Avinash Singh

Delivered conversational commerce agent systems for Resolve AI, supporting deployment across a range of real enterprise use cases including Nasdaq-listed Resolve AI contexts. Built agent workflows integrating planning, orchestration, and LLM task execution with evaluation and governance loops.

Built the LangGraph orchestration for a multimodal shopping agent, combining a frontier planner and proprietary post-trained task models running in parallel, with cost/latency-aware routing to specialized sub-agents. Designed a hybrid agent memory strategy (short-term session state, long-term shop preferences, and episodic memory) using hybrid BM25+dense retrieval with sharpening search for planner and orchestration agents. Post-trained open-source task models from filtered tool-call trajectories using schema and execution checks, scoring against frontier planning on held-out sets. Implemented a production-to-training loop by turning rubric-scored reviews into regression cases with fine-tuning and gated evals. Led two-person voice AI evaluations comparing half- and full-duplex speech-to-speech models against an in-house cascaded voice agent (Pipecat) using tool-call accuracy and perceived latency benchmarks.

Forecasting: built an AWS SageMaker framework across weather, claims, sales, and operations data and implemented custom Stats Forecast models; achieved significant speedups and reduced MAPE below target thresholds, with deployments in the US and Germany. Marketing mix modeling: built ElasticNet, Ridge, and Lasso models attributing sales by channel to set marketing spend allocations.