"Knight Eterniti team on AMD AI League Hackathon"

Knight Eterniti is a solo entry for the Lablab x AMD AI Academy Challenge, built and tested on an AMD Instinct MI300X with ROCm 10.0. Both images are built on the mandated rocm/pytorch base with the weights baked in, and run with no network. Each keeps a resident model server behind a unix socket, so every grader call is a standard-library client that starts in milliseconds and always writes a valid JSON file. Mini-Challenge 2 (OCR): Qwen3-VL-8B-Instruct (BF16) reads US and Chinese licence plates and road signs from PNG, JPEG and TIFF. The prompt and deterministic post-processing apply the transcription rules: drop state banners and slogans, keep Chinese province prefixes, join multi-line signs top to bottom, add no units. 10/10 public samples, 0.3-0.5 s per image, 18.8 GB VRAM, 45.7 GiB image. Mini-Challenge 3 (RAG): answers questions over a mixed corpus and cites the exact files each answer needs. Each file is parsed in an isolated worker process (PDF, DOCX with tables, multi-sheet XLSX, CSV, logs, code), so unreadable, encrypted or unknown files are skipped without stopping the walk. Images and scanned pages are transcribed by the same Qwen3-VL model. Withdrawn and superseded revisions are kept out of the model's context. Retrieval fuses identifier-aware BM25 with bge-small embeddings, and for multi-hop questions it follows rare identifiers such as ticket numbers and error codes across files. Citations are chosen by necessity: a file is kept only if it holds the value or supplied the identifier that leads to it, and an answer found in no document is refused. 10/10 public samples (200/200) with exact citation sets, 1-2 s per question, about 12 s startup, 20 GB VRAM, 45.8 GiB image.