Building LLM agents that act in the real world.
NUS Computer Science · Singapore
| Project | What it does |
|---|---|
| AEGIS: LLM attack-path agents · paper (arXiv) | Multi-agent system for the Singapore Armed Forces that discovers and validates exploit chains with LLMs. Runs 100+ step attack paths. |
| AI-safety classifier · model (Hugging Face) | Open-source harmful-content classifier, 220,000+ downloads. |
| Aggregate · site | Startup on a $25K AWS grant: browser agents collecting from 50+ sources daily (~98% success) into personalized digests. |
| Straits Cipher: historical archives RAG · demo video · repo | Citation-backed Q&A over ~7,000 colonial archive pages: OCR + hybrid-search RAG, ingested in ~30 min at $0.82 per 1,000 pages with parallel vLLM on burstable GPUs. Team of 3. |
Notes on agents, building and random things → rawthoughts.blog