Watch the workflow run, then make it work for you
Thirty to forty-five minutes of repetitive preparation per product becomes five to ten minutes of review. Nothing about the pipeline is fixed — the models, the thresholds and the governance rules are all yours to set.
Secure
Vendors never touch your core systems. Veranda sits between them, validates every record, and logs each decision. Sensitive data stays in your environment.
Fast
A thousand-variant batch moves from supplier feed to channel-ready record in minutes. The boring, repetitive preparation is already done before review.
Adaptable
Every step is configurable — field mapping, taxonomy, margin floors, and routing rules. Change them with data, not a release.
Works with the latest AI models
Each pipeline stage calls a model slot, not a vendor. Plug in frontier models, your own fine-tune, or a rules-only fallback as each category needs.
Five steps, each one configurable
Step 1
Submit
CSV, EDI or API. The vendor keeps their own format; the mapper handles the translation.
- ·Field mapping per vendor
- ·Required column set
- ·Batch size limits
Step 2
Enrich
Copy, attributes and imagery filled in from the record and the vendor's source material.
- ·Model choice per field
- ·Tone and length rules
- ·Do-not-invent field list
Step 3
Structure
Normalized to your taxonomy, units, colorways and channel contracts.
- ·Taxonomy version
- ·Unit and size grammar
- ·Channel field maps
Step 4
Govern
Margin floors, claim screening and completeness scoring decide publish, correct or hold.
- ·Margin floor by category
- ·Prohibited claim lists
- ·Confidence threshold to hold
Step 5
Act
Clean records publish to the channels. Questionable records land in a review queue with a reason.
- ·Auto-publish scope
- ·Reviewer routing
- ·Rollback window
Bring your own models
Each stage calls a model slot, not a vendor. Swap the model behind a slot, run both against the same batch, and compare the scores before you commit.
Copywriting
Today: Frontier general model
Swap for: Smaller in-house model, or your own fine-tune on approved brand copy
Attribute extraction
Today: Structured-output model with strict schema
Swap for: Rules-only mode for categories where invention is unacceptable
Image understanding
Today: Multimodal model for colorway and material checks
Swap for: Vendor-supplied attributes trusted, model used only as a second opinion
Claim screening
Today: Classifier plus deterministic term lists
Swap for: Your legal team's list, loaded as policy and versioned
Governance rules you write
Rules are data, not code. Edit them, version them, and see every decision they made in the governance log.
Margin floor
Set per category — 40% on Decor, 32% on Heating. Anything under stops with the arithmetic attached.
Claims policy
Medical, safety and performance claims are screened against a list you own and can edit without a release.
Completeness score
Decide the score at which a record can publish itself, and which missing fields are fatal versus cosmetic.
Human routing
Route by vendor, category, dollar exposure or confidence. New vendors can start at 100% review and taper.
The promise is review, not replacement
The ecommerce manager stays. What disappears is the repetitive preparation. Save 25 minutes across 1,000 products and that is roughly 417 labor hours returned to merchandising judgment.