30 minutes become 2 — and the model never chooses where to cut

🎯 The model scores, it doesn’t cut

It sees [042] 4.2s talking_head "texto" and returns a score. Timing comes from the word-level timestamped transcription, so the cut boundary is a real word boundary — never an invented timestamp.

🔁 Deterministic and inexpensive to redo

The selection (backpack by score, quota per topic, hook and closing anchors, coherence) is code. Even notas.json, even plan.json. Editing the plan by hand and rendering again costs zero model calls.

🔌 Provider can be switched by phase

Transcription with Groq or local Whisper; scoring with GLM, Gemini, Ollama, or Claude Code itself; TTS with inemavox or ElevenLabs; image generation with flux-2-klein. One line in the config.yaml or a flag.

Mode A — condensed

The default. Keeps the presenter and cuts only what works. Variant A+ (--substituir gerado) replaces presenter segments with generated illustrations and preserves the original audio.

Mode B — no presenter

Keeps only slides, screen demos, and charts; the model writes a script in Brazilian Portuguese, narrated with TTS, while the original audio becomes a bed at −18 dB.

Mode C — visuals only

Same selection as B, without the narration layer — for when the visual material already explains itself.

ffmpeg and ffprobe

Handle all the cutting, audio mixing, and headline. Rendering runs inside a scope with a memory limit.

# Debian/Ubuntu
sudo apt install ffmpeg

Python dependencies

yt-dlp to download, PySceneDetect to find the cuts, mediapipe and OpenCV to detect faces.

# in the project root
pip install -r requirements.txt

API keys

Read at runtime from ~/projetos/openpcbotv2/.env e ~/projetos/wifi/.env. No keys are copied into the project.

# what is used
GROQ_API_KEY
OPENROUTER_API_KEY
FAL_KEY  # only in mode A+