Dali Guo

I use Chronos-2 as a foundation model for demand forecasting, with cross-learning to draw on patterns shared across related time series. Context engineering is a central part of this work: dynamic-granularity batching controls how information is grouped and presented to the models, making context construction part of the forecasting design.

An agent layer built with Strands Agents SDK and AWS Bedrock brings these forecasts together with statistical signals and changing business context to propose forecast adjustments. This connects demand patterns learned by the foundation model with contextual information that can inform forecast calibration.

The system couples generative reasoning with a deterministic evaluation harness, which checks proposed adjustments against realized demand. That feedback drives programmatic optimization in DSPy, refining the agents’ reasoning and operating thresholds as demand outcomes become available.

Delivery-aware recommendations need to account for product complementarity, personal preferences, and fulfillment constraints. I developed a generative retrieval model that uses an RQ-VAE to turn graph embeddings into semantic product IDs, then conditions a Transformer on user embeddings through soft prompts. Context dropout encourages the model to generate complementary suggestions.

To support real-time recommendations, I decoupled offline personalization from nearline delivery signals and implemented constrained decoding with a trie to enforce delivery eligibility. I evaluated relevance and diversity through backtesting, with a separate check at serving time to confirm that suggested products could ship with the existing order.

My work in network science examines a limitation of fitting generative models to individual observations: a model can reproduce one network yet miss the variation across a population. I developed statistical comparisons and entropy measures to evaluate that variation in observed and generated network collections, making population variability an explicit part of model assessment.

Alongside this evaluation work, I implemented generative network models using VAEs in PyTorch and analyzed the relationship between functional brain connectivity and behavior. Read the network variability study (opens in a new tab).