.png&w=256&q=75)
🤓 Latest Submissions
KnowRisk
The problem. Modern semiconductors depend on a fragile, highly concentrated supply chain. Around 90% of advanced fabrication sits in Taiwan, EUV lithography comes from a single supplier in the Netherlands, HBM memory is concentrated in South Korea, and China gates key materials. Single-source dependencies, export controls, and multi-month lead times create systemic risk that procurement and risk teams struggle to see, quantify, or act on. What KnowRisk does. KnowRisk turns that tangled network into decisions. First, it ranks: a machine-learning model trained on a real 180,519-row supply-chain dataset scores every component and flags the critical top decile. Second, it explains: a locally hosted LLM converts each score into a structured brief covering the primary risk factor, a disruption scenario, and an actionable mitigation, plus a news-aware Q&A chat. Third, it maps: the dependency network is shown as a tiered graph and an interactive 3D globe that draws risk-colored arcs between countries, tracing a component's full multi-tier supply chain across the world. How it works. A FastAPI backend serves a single-screen dashboard with a ranked risk list, AI brief cards, graph and globe views, and analyst chat. It combines a multi-tier supplier graph, a scikit-learn GradientBoosting classifier, and an LLM agent that blends the model's disruption probability with a topological risk score. Powered by AMD. KnowRisk is local-first and has no proprietary cloud LLM. The full AI stack runs on AMD through ROCm: the model loads onto an AMD Instinct MI300X with PyTorch, GPU status is reported live via rocm-smi at every stage, and each answer surfaces its inference chip and latency in the UI. This makes AMD compute usage explicit, verifiable, and central to the product.
13 Jul 2026
