SJTU-> USC @ FORTIS Lab LLM Reasoning | Agentic System | World Model
Aegis. Runtime policy enforcement for AI agents. Cryptographic audit trail, human-in-the-loop approvals, kill switch. Zero code changes.
370Sovereign-OS. Constitution-first AI orchestration: one Charter (YAML) defines mission, budget & rules. CEO plans → CFO approves → Ledger tracks every cent & token → Auditor scores. 16 workers, Stripe, MCP. Think. Audit. Execute.
101Lithos.AI. A modern SaaS platform for renewable material construction management. Features AI-powered blueprint generation, real-time factory monitoring, construction progress tracking, and environmental impact analysis. Built with React 19, TypeScript, and Tailwind CSS.
10Telegrambot. ⚽ A viral Telegram football roast bot for the 2026 World Cup. Generate savage roast images, fake betting slips, and absurd VAR decisions using AI.
9BrewNet. Swift
9Topify-GEOAGENT. Python
7AIlearn. TypeScript
5Easylearningsniper. TypeScript
5Dynabridge-Discovery. Python
5VLMPretrain. Python
5cartpole. Jupyter Notebook
5sokoban. C++
4Shortest-Path. C++
4Game.
4Worldcup. TypeScript
4Verifiable_agent. Python
3parallel-hybrid-retrieval. Parallelizing a hybrid sparse-dense retrieval pipeline for long-term AI agent memory with OpenMP. Features BM25 + HNSW fusion, SIMD/MaxScore optimizations, temporal partitioning (47x speedup), and three-tier hierarchical memory.
3RecallOps. TypeScript
3thought-kv-hierarchy. Semantics-aware memory hierarchy for KV-cache in LLM reasoning. Offloads low-importance reasoning tokens from HBM to DDR instead of evicting them, preserving accuracy at reduced memory cost.
3federated-agent-audit. Privacy audit for multi-agent AI systems (CrewAI · LangGraph · AutoGen) — detect compositional data leaks without the central auditor ever seeing raw content.
3Chronoception-agent-temporal-cognition. Diagnosing and fixing temporal cognition failures in LLM agents — the Augustine Problem of agents that speak fluently about time without knowing what it is. Hosts ChronoBench (diagnostic benchmark across 25+ models) and ChronoStack (training framework for wall-clock grounding).
3sparse-index-tracking. Python
2Visteon-CabinVLMBenchmark. Python
1