What "substantially reduces manual tagging" actually looks like in production
This demo shows the exact two-pass tagging architecture proposed in the response to your RFP. Pick a sample asset, run the pipeline, and watch how Claude classifies against the 20-field taxonomy, validates its own confidence, runs reading-level assessment on family-facing content, and routes ambiguous tags to a human review queue. No LLM call is made; this runs the same logic against pre-computed model outputs to demonstrate the architecture without API costs.
Two-pass classification. Constrained enums. Confidence scoring.Reading-level gate. Family-Accessible / Plain Language thresholds.HITL queue. Below-threshold tags route to review.
1. Pick a sample asset to tag
Step 1 of 3
Select a sample asset above to preview.
Idle
2. Pipeline execution log
Step 2 of 3
Run the pipeline to see live output.
3. Result: Airtable taxonomy fields
Step 3 of 3
Tagged fields with confidence scores will appear here.
Human-in-the-loop review queue
After pipeline completes
—
Auto-tagged (high confidence)
—
Routed to review (low confidence)
—
Status blocked (reading level fail)
Queue items will appear here after pipeline runs.
How this works in production
This demo runs in your browser with no LLM calls. In your production deployment, the same logic runs through Make.com calling the Anthropic API. Asset submission triggers the pipeline:
New asset submitted in AO Docs → webhook fires to Make.com with asset metadata and content text.
Pass 1: Classification — Claude tags against the 20-field schema with each field's enum constrained in the response schema. response_format: json_schema eliminates freeform value drift.
Pass 2: Self-validation — Claude reviews its own classification, scoring confidence per field. Fields below the configured threshold (default 0.75) flag for human review.
Reading-level gate — Family-facing content runs through a separate prompt scoring against Family-Accessible (6th grade) and Plain Language (4th grade). Failures block Active status.
Companion content suggestions — semantic similarity surfaces likely companion assets. These are suggestions for human approval, never automatic links.
Write to Airtable — high-confidence fields populate directly. Low-confidence fields write with a Flag for Review status. Reading-level failures block Status = Active.
The accuracy validation against your sample of 50 already-tagged assets establishes the baseline reduction in manual tagging burden. That number — measured in hours per asset before vs. after — is delivered as part of the Phase 3 sign-off package.
Built by ResultantAI · Architecture proof for Acelero / Shine Early Learning · April 2026