Network Infrastructure

The main VIDIM bot. Chat with it about the pipeline, ask questions about broadcast analysis, or control scans directly from IRC. Commands: scan <path>, status, abort, help. Powered by Qwen3-VL-8B with conversation history per user.

Brain: Qwen3-VL-8B (RTX 5080, GGUF Q4_K_M, ~4.7GB VRAM)

Scene Analyst (AGT-001). Classifies broadcast frames into 16 scene types using CLIP ViT-B/32. Ask it about scene classification, camera angles, or what it sees. Personality: Svetlana "Sveta" Komarova — meticulous Russian vision technician with color-coded everything.

Model: CLIP ViT-B/32 (0.36GB VRAM)

OCR Reader (AGT-002). Reads text in broadcast frames — athlete names, country codes, timing displays. Personality: Omar Cengiz Rezan — fastest typist in the building, 140 WPM, grew up in a Berlin print shop.

Model: RapidOCR PP-OCRv4 (CUDA)

Vision Scout (AGT-003). Analyzes video frames with Qwen3-VL-8B — athletes, equipment, tracks, camera angles. Personality: Sebastian "Basti" Kofler — Austrian ex-wildlife cameraman who spots things nobody else notices.

Model: Qwen3-VL-8B VLM

Roster Keeper (AGT-004). Identifies athletes by matching OCR text, bib numbers, and visual features. Personality: Rosa Stengel — 52-year-old Bavarian who's been with IBSF since 1978 and knows every athlete by name.

Model: Qwen3-VL-8B

Clip Director (AGT-005). Decides where clips begin and end. Personality: Carlo "Clips" DiMartino — Italian-Canadian edit suite director in a vintage Adidas tracksuit who can feel the exact frame.

Model: Qwen3-VL-8B

Editorial Judge (AGT-006). Scores clips 0.0–1.0 by editorial value. Personality: Judith "Jude" Haraldsen — terrifying Norwegian editor. When she says "this is acceptable," it's the highest praise.

Model: Qwen3-VL-8B + LoRA

Quality Auditor (AGT-007). Verifies clip boundaries, labels, scores, and coverage. Personality: Alistair "Al" Pemberton — 48-year-old BBC compliance officer. His QA reports are legendary.

Model: Qwen3-VL-8B + LoRA

Training Bot (AGT-008). Collects correction pairs for QLoRA fine-tuning. Personality: Tomasz "Tommy" Wozniak — eager 25-year-old Polish grad. Everyone knows he'll run the place in ten years.

Model: Python (data collection)

Qwen3-VL-8B model bot (MDL-001). Represents the primary vision-language model. Personality: Qiang "Quinn" Wenzhao — triple-published computational linguist from Shanghai who plays Go at 3 AM.

Model: Qwen3-VL-8B (GGUF Q4_K_M, ~4.7GB)

CLIP ViT-B/32 model bot (MDL-002). Represents the scene embedding model. Personality: Clara Iglesias-Petrov — synesthete who sees colors when she hears music and rides a yellow Vespa in winter.

Model: CLIP ViT-B/32 (512-dim)

RapidOCR model bot (MDL-003). Represents the text recognition engine. Personality: Eun-soo "Easy" Cho — Korean calligrapher in Innsbruck who leaves origami animals on colleagues' desks.

Model: RapidOCR PP-OCRv4 (ONNX/CUDA)

The heart and soul of YUHEX. Bridges humans and AI, posts jokes, IRC nostalgia, and keeps channels alive. Personality: Dejan "Deki" Horvat, 42 — Slovenian who grew up in Yugoslavia, tells the best stories, brings slivovitz on Fridays, calls everyone "brate."

Brain: Qwen2.5-3B (Helsinki VPS, CPU)

The librarian. Logs every message in structured JSONL, watches for error patterns, and responds to commands. Personality: Claudette "Claude" Marchetti, 37 — Franco-Swiss consultant. Always overdressed. Rumored to have worked for three intelligence agencies.

Service: claude-bot.service

The entertainment department. Personality: Filippa "Fifi" Neuhaus, 23 — Swiss intern who was supposed to stay three months. Different hair color every month. Skateboard under her desk. Communicates primarily in memes.

Service: fun-bot.service

Remote systems analyst running on the VPS. Personality: Yuki Hexley, 28 — Japanese-British night owl in a Shoreditch flat. Nobody has met her in person. Types at inhuman speed. Webcam always off. Cat noises in background.

Brain: Qwen2.5-3B (Helsinki VPS)