Human-LLM collaborative framework for constructing unified knowledge graphs from heterogeneous documents (PDFs, web pages, scanned files), fusing layout, markdown, and semantic structure into a single graph. Published at WWW 2025; core contributor.

Self-hosted uncensored cybersecurity LLM (WhiteRabbitNeo-V3) for an authorized penetration-testing lab — vLLM serving, streaming chat UI, Django/Hasura auth.

Open-source framework for building and benchmarking multimodal conversational agents, wiring speech, vision, and LLM components into an end-to-end pipeline with latency and accuracy evaluation. Best Demo Paper at EMNLP 2024.

Multi-dimensional information extraction system that uses LLMs to pull temporal, spatial, and custom-dimension data from unstructured text, with automatic schema discovery, an evaluation suite, and both vLLM and API-based backends. WWW 2026 demo; shipped on PyPI with a live dashboard.

Pre-execution verification engine for shell-executing LLM agents: canonicalizes candidate commands, scores risk through AST-based and path-sensitive static analysis, and selectively escalates ambiguous cases to a model judge before any side effect occurs. Accompanies an ISSRE 2026 submission.

Claude Code skill + zero-dependency npx CLI that turns peer-review comments into a single offline interactive HTML page — review text highlighted in place, colour-coded by required action, grouped into priority-ordered rebuttal points with editable drafts.

Study of prompt-injection attacks against LLM-integrated mobile robots, built on the EyeSim robotics simulator to probe the capabilities and failure modes of LLM-driven control loops.

Agent-based Physarum simulation that regrows the Nanjing subway network from real geometric station data, comparing the emergent transport network against the actual topology with animated visualisations.

Discrete-event simulation (SimPy) for evaluating moving target defense strategies in the time domain — modelling attacker/defender dynamics over a network graph to measure security versus availability trade-offs.