Our work spans three connected frontiers, all asking how data should be managed when AI is the primary consumer.
We bring data integration and machine learning together — so heterogeneous data and rich model zoos meet in one lake, and serving models becomes a query.
Model Lake
Amalur explores the convergence of data integration and machine learning — automating how scattered training data across silos is integrated for downstream models. It is the foundation of the group’s Model Lake vision, where heterogeneous data and rich model zoos meet in one place.
LLM serving
TranSQL serves large language models with relational queries — compiling model inference to SQL so that LLMs can run inside a database engine, even on low-resource hardware.
We train models and generate synthetic data across organisational silos, extracting value from data without ever moving the raw records.
Synthetic data
SiloFuse generates cross-silo synthetic tabular data using latent diffusion models, so organisations can share realistic data without ever exposing raw, feature-partitioned records.
Time series
WaveStitch performs flexible and fast conditional time-series generation with diffusion models, stitching together realistic signals under user-specified constraints.
We rethink classic data-management problems — query optimisation, entity matching, anomaly detection — for the NISQ-era quantum processor.