Drop in any talking-head video. Our pipeline transcribes it, finds the logical fallacies in what's actually said, and splices in picture-in-picture explanations — a publish-ready vertical remix, for about five cents.
Explainer and "this argument is wrong because…" content performs — but it's slow, manual, and expensive to produce. Remix automates the whole loop: speech in, a finished, captioned, voiced fallacy-breakdown reel out. Same recognisable host voice every time, every claim backed by the speaker's own words.
Each stage writes a JSON artefact, so any step can be re-run or swapped. Because stages run sequentially, peak memory is a single model — the whole thing fits on a 16 GB laptop and scales the same way in the cloud.
Auto-labelling fallacies on real people is reputationally dangerous. The defensibility is in the guardrails that keep it honest and cheap.
A fallacy is discarded unless its quote is found verbatim in the word-level transcript. Every card points at words the person actually said, with a timestamp.
Each candidate is independently re-verified in context and dropped unless confirmed. Measured: 7/8 kept on clear arguments, 4/4 correctly rejected on casual chatter.
The LLM emits schema-constrained JSON (GBNF grammar), so detections parse deterministically — no brittle prompt-scraping, no malformed runs.
Open models (Whisper, Qwen, Kokoro) on commodity GPUs — no per-call vendor lock-in, a locked brand voice, and customer footage that never has to leave our infrastructure.
Compute is a rounding error. One ~3-minute reel — transcription, ~20 LLM calls, voiceover, and render:
| Component | Approach | $/video |
|---|---|---|
| Transcription | Whisper (hosted/own) | 0.002 |
| Reasoning (~20 LLM calls) | 7–8B model | 0.001–0.005 |
| Voiceover | Kokoro / budget TTS | 0.00–0.007 |
| Render | ffmpeg CPU | 0.005–0.02 |
| GPU wrap (self-host) | serverless L4, ~4 min | ~0.05 |
| All-in COGS | managed → self-host | 0.03–0.12 |
Three fulfilment models; the cheapest flips as volume steadies. Monthly compute, budget voice.
| Reels / month | Managed APIs | Serverless GPU | Dedicated GPU |
|---|---|---|---|
| 500 | ~$15 | ~$30 | $216 |
| 5,000 | ~$150 | ~$300 | ~$216 |
| 50,000 | ~$1,500 | ~$3,000 | ~$650–1,080 |
Start on pay-as-you-go to launch with zero idle cost; switch to owned GPUs past ~3–5k reels/month, where utilisation makes them 5–10× cheaper per unit. Same codebase either way — only a config profile changes.
Not a concept deck — a working, staged pipeline with real outputs on real clips.
Ingest → ASR → argument mining → fallacy detection → scripting → voice → splice/render, one re-runnable artefact per stage.
Quote-back guard and self-verification live, with a full audit trail of what was flagged and what was rejected and why.
Async job service (upload/URL → webhook → signed download). MVP on managed APIs + R2; GPU-owned path ready for scale.
Direct URL ingest and a dashboard; rights controls and rate limits built in from day one.