- Recently
- MAFBench accepted at SIGMOD 2027
- Demonstration Chair of VLDB 2027
- Three PhD students and one MSc student join CoDS
What we work on
Agents have no governed way to accumulate what an organization knows, and no way to establish that what they accumulated is trustworthy. We work on this at the data layer, and release our systems open source through CoDS-GCS.
- Governed memory and tacit knowledgeAgent memory as a governed state machine: typed dependencies, semantic revision, declarative forgetting, multi-tenant privacy.
- Action contracts and governed mediationStructural enforcement of what an agent may read, write and promote; addresses belief laundering.
- Validation of agent-generated codeAssertion-based and derivation-aware validation that replaces probabilistic self-judgment.
- Agent security and threat detectionMulti-agent attack surfaces, memory poisoning and prompt injection. Builds on OCR-APT and MEGR-APT.
- Graph foundation models for agentic reasoningMotif-centric inference over enterprise knowledge graphs, giving agents structured recall at scale.
Recent systems
- SIGMOD 2027MAFBench, an evaluation of nine multi-agent LLM frameworks: five architectural dimensions, six design principles.
- SIGMOD 2026Chatty-KG, multi-agent conversational question answering over knowledge graphs; 39 to 80 F1 with the same LLM.
- VLDB 2026ReSequel, LLM-assisted SQL rewriting, up to 16× over native engines.
- CCS 2025OCR-APT, advanced persistent threat stories reconstructed from audit logs.
- EMNLP 2025RAGvis, knowledge-grounded RAG for exploratory data analysis and visualization.
- VLDB 2025CatDB, catalog-guided, LLM-based generation of data-centric ML pipelines.
KGLiDS was adopted by Google and Kaggle teams, and RAGvis is released in Google’s own GitHub organization.