Runflow
Introducing Sentinel

Your AI pipeline shipsbad images.Sentinel stops that.

Automated quality evaluation built into every AI Image and Video worfklow. Detect, score, and filter AI-generated images before they reach your users.

Get beta accessSee how it works
The Problem

You're generating at scale.Nobody's checking the output.

When AI pipelines run at volume, defects multiply. No team reviews 10,000 images a day. Without automated evaluation, broken images reach your users.

Invisible

You can't see what's broken

Face distortions, wrong backgrounds, skin tone issues, logo placement errors. At API scale, defects are invisible until a user complains.

Manual & slow

Human review doesn't scale

Manual QA works for dozens of images. Not for thousands. The moment you scale, quality control becomes the bottleneck — or disappears entirely.

Costly

Bad outputs cost real money

Every image that fails QA after delivery is a refund, a complaint, or a churn event. Defect cost doesn't show in COGS — it shows in your NRR.

How it works

Score every image.Deliver only the best.

Sentinel scores every image your pipeline generates. Passing images flow to delivery. Failing images get a recommended action. Available inside Runflow pipelines and as a standalone API.

01

Image enters evaluation

Sentinel intercepts the output from any generation node. The image is analyzed using a vision model trained on quality signals specific to your use case — not generic AI evaluation.

02

Multi-dimensional scoring

03

Sentinel flags and recommends

sentinel · evaluation result
pipeline: headshot_v2 · output_38291.jpg
Face detected
Likeness match
91
Background clean
88
Smile detected
95
Skin tone natural
84
PASS — Score 91.4
Image cleared for delivery
Evaluation layers

Built for your use case,not the generic one.

Sentinel evaluates at three levels. From universal quality checks to custom business rules you define. Every layer is configurable.

Layer 01

Generic Quality

Universal checks that apply to any AI image generation. Runs automatically on every pipeline, zero configuration.

Prompt-to-image alignment
Artifact detection
Composition scoring
Sharpness & exposure
Most used
Layer 02

Use Case Specific

Pre-tuned criteria for Runflow's production pipelines. Headshots, fashion, on-model — each with its own quality standard.

Face fidelity & likeness
Garment & logo accuracy
Background consistency
Expression & skin tone
Layer 03

Your Custom Rules

Define what matters to your business. Sentinel checks your edge cases on every single run, at any volume.

Custom quality schema
Business-specific criteria
Threshold configuration
Team-shareable templates
Proof

Deployed in production.Results are real.

Case study — BetterPic.io

Generate 4× more candidates.Deliver only the best ones.

BetterPic runs their headshot pipeline at 4× volume. Sentinel scores every output in real time, then selects only the highest-performing images to deliver to users.

Users receive a tightly curated set of top headshots — matching expression, likeness, background, and skin tone criteria. The rest are suppressed automatically. No human reviewer involved.

Zero manual review. Zero bad images delivered.
📥
User uploads photos
Input: 10–20 reference photos
Runflow generates at 4× volume
e.g. 240 candidate images per user
🎯
Sentinel scores every image
Face · expression · background · skin tone
Top 60 delivered to user
Highest-scoring only · zero manual review

"We generate 4× more candidates and let Sentinel pick the best ones. Users get a curated set of top-quality headshots — without us reviewing a single image."

Thibaut Hennau, CEO — BetterPic.io
Generation volume
Candidates per user
Images delivered to user
Top-scored by Sentinel
AllBest 60
Manual QA headcount
RequiredZero
What changed for the user
Before Sentinel

User receives all generated images — mixed quality, unscored. Some are great. Many have expression issues, background artifacts, or likeness mismatches. User must pick manually.

After Sentinel

User receives 60 curated images — all scored above threshold on face, expression, background, and skin tone. More candidates were generated, only the best surface.

Note: Exact generation ratio and delivery count to be confirmed with Ziad before publishing. Numbers reflect current production behaviour.
Developer-first

One call to add quality control to any pipeline.

Sentinel is available as a standalone API and built into every Runflow managed pipeline. Add it in minutes. Configure it for your use case. Never touch it again.

Instant integration

Plug Sentinel into any existing Runflow pipeline with a single parameter. No pipeline rebuild required.

🎯

Configurable thresholds

Set pass/fail scores per dimension. Sentinel adapts to your quality bar — strict for luxury, lenient for drafts.

🔁SOON

Auto-fix executionComing soon

Today Sentinel flags failing images and recommends the right action: re-generate, targeted auto-fix, or manual review. Automated execution of fix workflows is on the roadmap.

POST /v1/pipelines/headshot
{
  "input_images": ["photo_001.jpg"],
  "style": "corporate_headshot",

  "sentinel": {
    "enabled": true,
    "threshold": 80,
    "on_fail": "suppress",
    "top_n": 60,
    "criteria": [
      "face_likeness",
      "smile_detected",
      "background_clean",
      "skin_tone_natural"
    ]
  }
}

// Response
{
  "delivered": 60,
  "suppressed": 180,
  "avg_score": 91.4,
  "output_urls": ["cdn.runflow.io/..."]
}
Beta access

Stop shipping bad images.Start with Sentinel.

Quality control is not optional at scale. It's what separates an AI product from a reliable AI product.

Managed Pipelines

Sentinel included

Already on a Runflow managed pipeline? Enable Sentinel with one parameter. Quality control on every run.

  • Use case-specific evaluation presets
  • Score reporting per image
  • Auto-fix execution (coming soon)
Get beta access
Custom Pipelines

Sentinel configured

Building your own pipeline? We configure Sentinel for your exact use case, quality criteria, and thresholds.

  • Custom evaluation schema
  • Your criteria, your thresholds
  • Shareable team templates
  • Dedicated onboarding call
Get beta access

No commitment required. We onboard 5–10 teams per month.