Agentic Product ML
Building practical assistant experiences that connect user intent, business context, and reliable actions.
Agentic AI / Probabilistic ML
Vincent Zheng
Applied ML Scientist at OpenTable working on Concierge agent experiences.
I focus on applied AI systems that turn user intent and contextual signals into useful product experiences. My current work is on Concierge agent experiences at OpenTable. My research background spans LLMs, retrieval, ranking, probabilistic ML, and foundation models.
Focus
Current Themes
Retrieval and Ranking
Using search, embeddings, ranking, and evaluation to improve decision support in real products.
Probabilistic ML
Research background in uncertainty, structured prediction, and time-series foundation models.
Selected Work
Publications
Overcoming the Modality Gap in Context-Aided Forecasting
Introduces CAF-7M and DoubleCast for context-aided probabilistic forecasting with textual information.
Probabilistic Traffic Forecasting with Dynamic Regression
Extends error-correlation modeling to spatiotemporal forecasting with a matrix-variate autoregressive process and non-isotropic training loss.
Multivariate Probabilistic Time Series Forecasting with Correlated Errors
Introduces an efficient parameterization of cross-covariance matrices for multivariate probabilistic forecasting.
Better Batch for Deep Probabilistic Time Series Forecasting
Uses generalized least squares in the temporal domain to account for autocorrelated errors in deep probabilistic forecasting.
Projects
Applied AI and Research
OpenTable Concierge Agent
Applied ML for concierge-style dining assistance, connecting user intent, context, ranking, and reliable product actions.
DoubleCast
Context-aware prediction with CAF-7M, Chronos, Qwen3-14B, and DualT5 cross-attention.
Dynamic Regression
Structured spatiotemporal prediction with dynamic error-correlation modeling.
Background
Recent Roles
Applied ML Scientist, OpenTable
Working on Concierge agent experiences for dining discovery and assistance.
Visiting Researcher, ServiceNow Research
Worked on LLM and foundation models for context-aware forecasting.
PhD Researcher, McGill University
Developed probabilistic time-series methods for calibrated uncertainty estimation, including temporal autocorrelation, multivariate cross-covariance, and spatiotemporal error-correlation models.
Mitacs Accelerate Intern, ExPretio Technologies
Built adaptive multi-horizon models for uncertainty-aware demand modeling.
Recognition
Selected Milestones
- Mitacs Accelerate, ServiceNow Research, C$32,000, 2025
- FRQNT B2X Scholarship, Fonds de recherche du Quebec, C$77,000, 2022
- McGill Engineering Doctoral Award, C$111,000, 2021
- National Scholarship for Graduate Students, China, 2018