Skip to content

Stop posting AI slop.
Start posting work that survives a debate.

Built around a three-model editorial review. A Savitzky–Golay smoothed virtual cameraman re-films your 16:9 footage as cinematic 9:16. Your brand voice survives. Your audience notices.

LIVE · job_8f3a · clip 04/12
round 2/3
FACE · 0.97
consensus is the moat
Brand Guardian Claude Opus 4.7
Voice match: tight. No 'tapestry'.
PASS 9.4
Visual Analyst Gemini 2.5 Pro
Speaker tracked 0:00–0:42. No drift.
PASS 9.1
Retention Lead GPT-4o
Hook lands at 0:03 — trim 1.4s.
REVISE 7.8
CONSENSUS · WEIGHTED
8.77 / 10
ACTION
trim · re-vote
SKIP THE
manual edit
on every pillar video
Models in the board
3 frontier-tier
Audit iterations
≤3 until consensus
API cost / job
$0.05 85% slashed in S7
Pipeline tests passing
168/168 S11 verified
THE EDITORIAL BOARD

A three-agent newsroom, debating every clip.

Most tools fire one LLM at your video and hope. We run a MassGen consensus loop. Each agent gets a different system prompt, a veto, and the right to call a Blocker on the others until the clip survives three rounds.

PASS
9.4
ROLE

Brand Guardian

Claude Opus 4.7

Zero tolerance for AI clichés. Audits the draft against the speaker's actual tone.

  • Flags 'tapestry', 'delve', 'unleash', 'unlock'
  • Compares output tone vs. source transcript
  • Demands unique insight, not Google-result paraphrase
PASS
9.1
ROLE

Visual Analyst

Gemini 2.5 Pro

Multimodal alignment. The text must agree with what's on screen.

  • Visual-text drift detection
  • Fact density vs. 1-hour raw video
  • No clipping mid-thought or mid-sentence
REVISE
7.8
ROLE

Retention Lead

GPT-4o

Performance director obsessed with the scroll-stop. Demands a hook before second 4.

  • Platform-native retention curve
  • Pattern-interrupts every 7 seconds
  • SEO-tagged for the host platform
PROOF OF WORK

Your client sees the full debate, not just the clip.

Every job ships with the round-by-round transcript. The agents argue, vote, rewrite, vote again. That transcript is the trust signal we sell.

brand_guardian Hook reads as generic LinkedIn-ese. Speaker said "consensus is the moat" — use his words.
retention_lead Agreed on the line. But cut 1.4s of preamble before it lands.
visual_analyst Face track stable 0:00–0:42. Crop holds. No drift.
brand_guardian Re-voting with the trimmed hook. PASS · 9.4
consensus reached · round 2/3 · 14.2s
THE PIPELINE

Six stages. Local-first. Stateless.

The math is the moat. Every stage is containerized and tested with golden-frame diffs. Migrating from Mac Studio to a GPU VPS is docker run + a DNS flip.

WALL-CLOCK
~98s
end-to-end, 60s source video
STAGE
01

Ingest

yt-dlp · uploads

Pull the source from YouTube, R2, or direct upload. Whisper large-v3 transcribes locally on Mac Studio or VPS.

≤2 min per hour
STAGE
02

Editorial Board

MassGen consensus

Claude / Gemini / GPT debate hooks, fact density, and brand voice. Slop filter strips the 25 banned terms.

3 agents · 3 rounds
STAGE
03

Vision pass

YOLO v8n → Gemini Flash

Hybrid vision finds the speaker every frame. Local YOLO handles 93% of jobs at $0. Flash catches the rest.

~$0.003/video
STAGE
04

Cinematic camera

SG smoothing + smootherstep

Savitzky–Golay removes jitter, smootherstep gives the pan weight, semantic zoom punches in on insight.

human-feel
STAGE
05

Render + grade

VideoToolbox / NVENC

HW-accelerated encode, animated ASS captions, 3 color-grade presets. Pillow PNG labels — no libfreetype.

≤90s per clip
STAGE
06

Quality gate

Gemini Flash review

Five-frame post-render review. Score 1–10. Soft gate at 6. Attestation signed and minted to the reputation graph.

score ≥ 7
BEFORE / AFTER

Most tools center-crop. We re-film.

Same 60-second source. Same speaker. Two pipelines. The left is what OpusClip ships. The right is what survives a three-agent debate plus a Savitzky–Golay smoothed virtual cameraman.

BEFORE · center-crop standard AI clip
⚠ drift
unlock your tapestry
5.2 /10
  • Hook lands at 0:11 — past the scroll
  • Caption hallucinates 'tapestry'
  • Speaker drifts off-frame at 0:23
AFTER · consensus-validated slop-busters
FACE · 0.97
consensus is the moat
9.1 /10
  • Hook at 0:02 — speaker's own words
  • Caption matches transcript exactly
  • Face tracked end-to-end
Pipeline run
~98s
end-to-end · 60s source
Audit iterations
≤3
until consensus
Cost per pillar video
$0.05
API · S7 reduction
PRICING

Managed, attested delivery. Not a seat.

Billed monthly. Cancel any time.

1 clip

Sample

Free

Free — one attested clip, plus the full multi-model debate transcript.

  • 1 cinematic 9:16 clip
  • Full agent debate transcript
  • Per-clip MassGen attestation
  • Email delivery in ~6 hrs
Request sample
Post-founding default

Standard

$499 / month

Standard — managed, attested delivery. Our post-founding default.

  • Managed, attested clip delivery
  • MassGen multi-model debate per clip
  • Brand-voice profile extraction
  • Private debate dashboard
  • Fair-use cap + metered overage
Custom

Agency

Custom

Agency / custom — dedicated throughput and GPU shard, priced to your volume. Talk to us.

  • Dedicated throughput
  • Dedicated GPU VPS shard (add-on)
  • Custom brand profile YAML
  • White-label sample pages
  • API access (Hono + Workers)
  • Priced per engagement
Talk to the operator

Pricing applies to live pipeline jobs. Sample requests are limited to one per workspace per quarter.

THE FAMILY

Plugs into the rest of Howlitzer.

Single sign-on across the constellation. Each consensus-validated clip mints a signed attestation into the reputation graph — so your Nexus profile carries proof of work, not just claims.

READY?

Your audience deserves better than a center-crop.
Give them the debate.

Submit one long-form video. Get peer-reviewed text clips, a cinematic 9:16 edit, and the full agent debate to prove the work — all under two minutes of wall-clock.