Systems and benchmarks
The CoDS-GCS GitHub organization open-sources more than 30 repositories. Each ships with reproducible benchmarks, documentation and an associated peer-reviewed publication.
External adoption. KGLiDS was adopted by Google and Kaggle teams for production data science workflows. Its follow-on system, RAGvis, is released inside Google's own GitHub organization (google/ragvis) and was published at EMNLP 2025.
Agentic AI systems and benchmarks (2025 to 2027)
-
MAFBench
Unified evaluation suite for multi-agent LLM frameworks, with an architectural taxonomy over five dimensions and a controlled study of nine frameworks. Yields six evidence-based design principles.
-
MemState
Prototype of Governed Evolving Memory on a property-graph backend, with topic-based storage, typed dependencies and declarative policies.
-
Chatty-KG
Multi-agent system for on-demand conversational question answering over large knowledge graphs. Doubles accuracy (39 to 80 F1) with the same LLM under a better architecture.
-
ReSequel
LLM-assisted SQL rewriting over PostgreSQL, MySQL and DuckDB: templatizes workloads, infers rules from the catalog, verifies candidates on downsampled data. Up to 16× over native engines and 22× over LLM baselines across eight benchmarks.
-
OCR-APT
Reconstructs advanced persistent threat stories from audit logs using subgraph anomaly detection and LLMs, with hard-gating assertion validation.
-
RAGvis
Knowledge-grounded RAG for exploratory data analysis and visualization; a reliable, cost-effective alternative to large-model EDA.
Released in Google's own GitHub organization.
-
CatDB
Data-catalog-guided, LLM-based generator of data-centric ML pipelines, with iterative self-correction validated on the actual data.
-
Chatty-Gen
Dialogue benchmark generation from knowledge graphs with retrieval-augmented LLMs, cutting cost from thousands of dollars to under one dollar.
Foundational systems
-
KGLiDS
Platform for semantic abstraction, linking and automation of data science over knowledge graphs. Underlies CatDB and RAGvis.
Adopted by Google and Kaggle teams for production data science workflows.
-
KGNet
GNN-enabled knowledge graph engine with task-oriented training and LLM-guided query-aware inference (KG-WISE).
-
KGQAn
Universal question answering over knowledge graphs, and GNN-based scalable AutoML with IBM Research.
-
KGpip
Scalable AutoML approach based on graph neural networks, developed with IBM Research.
I also contribute to aurum-datadiscovery (MIT DBG), and to E-Store (VLDB 2015) and Solid (WWW 2016, with Tim Berners-Lee).
Earlier research projects
These projects from before the current program keep their original pages:
- The Data Civilizer System: end-to-end data discovery, integration and cleaning in large enterprises.
- Managing Linked Data at Scale (Lusail): querying, integrating and sharing geo-distributed RDF graphs.
- Elastic in-memory OLTP (E-Store)
- Large-scale analytics on strings