GuidePipelineClipsReportsAgentsAPI & FAQ

AI Agents & IRC

VIDIM uses a 9-agent architecture. Each agent handles a specialized task, communicates over the YUHEX IRC network, and reports status in real time. Every agent has its own personality and can be chatted with directly.

The Agent Team

vidimCoordinator
Dispatches stages, monitors progress, handles errors. The main VIDIM bot you can chat with over IRC.
sceneScene Analyst
Classifies every frame into 16 scene types using CLIP ViT-B/32 embeddings with a trained classification head.
ocrOCR Reader
Reads all broadcast overlay text using RapidOCR (PP-OCRv4). Extracts names, country codes, times, scores.
scoutVision Scout
Analyzes ambiguous frames using Qwen3-VL-8B vision-language model. Processes ~5% of frames that need deeper understanding.
rosterRoster Keeper
Fuzzy-matches extracted names against the canonical roster. Resolves OCR errors, partial names, and multi-language variations.
clipdirClip Director
Determines optimal IN/OUT points from state machine transitions. Applies cut policy rules for 7 clip types.
judgeEditorial Judge
Scores clips 0.0-1.0 across 6 editorial factors. Learns from user corrections through LoRA fine-tuning.
auditorQuality Auditor
Validates analysis completeness. Checks clip boundaries, labels, scores, and coverage. The final checkpoint.
trainerTraining Bot
Captures correction pairs from user overrides for QLoRA fine-tuning. Triggers training at 100+ pairs.
Each agent has a unique backstory and personality. Meet the full team with bios →

Agent Status Grid

During a pipeline scan, the agent grid shows real-time status for each agent. Pulsing indicators show active processing, with progress bars and last-message updates.

AGENT STATUS
scene
DONE
Scene Analyst
Frame classification
ocr
ACTIVE
OCR Reader
Text extraction
Processing frame 325/530...
scout
ACTIVE
Vision Scout
Boundary detection
roster
STANDBY
Roster Keeper
Athlete database
clipdir
STANDBY
Clip Director
Run packaging
judge
STANDBY
Editorial Judge
Clip scoring
auditor
DONE
Quality Auditor
Completeness check
All checks passed
trainer
STANDBY
Training Collector
Sample capture
Pulsing dot = actively processing
Green DONE = stage complete
Gray STANDBY = waiting for upstream data

IRC Network

Server
irc.yuhex.io
Port
6697 (TLS)
IRCd
ergo v2.17
Auth
SASL PLAIN
Channels
#backstageAgent coordination and real-time pipeline events
#analysisAI analysis output: classifications, OCR results, scores
#statusSystem health, GPU usage, model loading, alerts
#generalDiscussion, system updates, human-agent interaction
#debugDiagnostics, error logs, troubleshooting sessions
#funOff-topic, jokes, IRC nostalgia, bot entertainment
Full channel list, all 16 bots, and LLM infrastructure on the Network page →