"PHAM MINH CANH team on AMD AI League Hackathon"

Most web agents hallucinate. This agent is built around one rule: never ship an unverified citation. Given a start URL and a natural-language request, it fetches the page, extracts readable text, and picks the windows most likely to contain the answer using IDF-weighted term scoring instead of blindly taking the first 6000 characters. It then asks a Token Lab LLM (DeepSeek-V4-Flash) for exactly one short verbatim quote. Every quote is checked against the fetched page text; if verification fails, the agent falls back to a guarded extractive answer and records fallback_reason instead of inventing evidence. When the page truly lacks the answer, it returns an honest not-found result. Results: 11/11 evaluation questions pass, 9/9 citations verify verbatim against the source page, 0 fabricated citations, 14 pages fetched, and a multi-page demo cites the correct page. A live Token Lab run also passes 11/11 (9 LLM answers, 2 guarded extractive fallbacks). 22 unit tests pass offline. The AMD stack is load-bearing: developed on ROCm 7.2 on an AMD Instinct MI300X VF GPU (wavefront 64) with HIP kernels for vector add, histogram and reduction, rocBLAS GEMM (73x over naive), a tiled LDS transpose (14x), and rocprofv3 counter profiling. The agent itself is pure Python standard library plus the AMD Token Lab API, so it runs with one command.