Tiffany Lighting Direct
Home and lighting retail
- To get a new supplier range live, down from weeks or months
- Days
- Extra hires needed. Cernel works like an extra employee on the team
- 0
Build versus buy
Build versus Buy for Product Data Enrichment
A demo that fills a few fields is quick to build. Running enrichment across a whole catalogue, in every language, every day, takes a team. A fair look at both paths for ecommerce, master data and IT teams.
- A demo
- Cernel at work
- A person decides
- Done
What building takes
A demo is easy. Running it is the work.
Building enrichment in-house can be the right call, and a first version comes together quickly. What decides the cost is everything after it: the parts that keep it accurate, affordable and running on every product, every day.
Getting data in and out
To build it in-houseConnectors for your PIM, ERP, shop and feeds, and a sync that writes the right value back to the right place without overwriting someone's edit.
With CernelShopify, Struct PIM and inriver sync both ways. CSV files, feeds and a REST API cover everything else. Integrations
Categories and attributes
To build it in-houseA model of your category tree, and Google's beside it, with rules that inherit down the tree, attributes that depend on each other and variants that stay consistent.
With CernelEvery product is placed in your categories and gets the fields that category needs, with rules set once high in the tree. Categories and fields
Agents per attribute
To build it in-houseOne big prompt per product goes generic. Good results take a prompt for every attribute in every category, tuned and kept up to date.
With CernelA specialised agent for each attribute, written in plain language by your team and tested on real products before it runs. AI agents
Research from supplier sources
To build it in-houseSearch and matching that finds the maker's own page for the right product, and a record of where every value came from.
With CernelAgents read the maker's and supplier's pages for the same product and keep the source with every value they write. Research agents
Quality and hallucination control
To build it in-houseChecks that catch broken characters, cut-off text and values that look right but are not. Plausible but wrong is the failure that costs most.
With CernelAnswers come from your own option lists, an agent that is not sure says so instead of guessing, and checks run on every output. Review steps
Translations
To build it in-houseEvery text in every market language, checked for tone, special characters and terms that must not be translated.
With CernelEvery market language in the same run, each with its own prompt, so copy reads native instead of word for word. Translations
Images
To build it in-houseA separate image pipeline, with its own models, prompts, file names and alt text.
With CernelMedia agents treat images as one more attribute: they work from a brief and the product's own data, and write the alt text and file names. Image generation
Review and approval
To build it in-houseA review tool for your team, rules for what needs a person, and a log of who approved what.
With CernelYou choose which values go live automatically and which get a review step. Review steps check every value and only ask a person when something fails or needs a decision, and every approval is logged. Review and approval
Evaluation and model upgrades
To build it in-houseA test set from real catalogues, judge models and the statistics to tell a real gain from noise, every time a new model comes out.
With CernelNew models are tested on real past work and rolled out once they win. Upgrades are part of the platform, not a project on your side.
Scale, monitoring and upkeep
To build it in-houseLong jobs that survive outages, retries that never pay twice, an AI bill that grows with the catalogue, a queue of change requests and the people who know how it all works.
With CernelCernel runs, monitors and tunes it every day. Your product team makes changes in plain language, without a ticket.
What Cernel already runs
One value. Everything behind it, already built.
Follow one field on one product. Every step is a part of the platform you would otherwise build, test and keep running yourself.
Your systems
The product arrives from your PIM, ERP, shop or a supplier file.
Categories and fields
It is placed in your category tree and gets that category's fields.
An agent per attribute
Each field has its own specialised agent, inherited from higher in the tree.
Research with citations
The agent reads the maker's own page and keeps the link.
Quality checks
The value must come from your option list and pass every check.
Translations
Every market language in the same run.
Images
A Media agent writes the alt text and file name.
Review and approval
A review step checks it and only asks a person when something fails or needs a decision.
Two-way sync
The approved value goes back where it belongs.
Example product. Names and values are invented.
For IT and engineering teams
A hard infrastructure problem and a hard AI problem, at the same time.
At catalogue scale, enrichment is a distributed system where every call has its own capacity, failure modes and price. This is the part a demo never shows.
Long jobs that survive anything
Jobs over a whole catalogue run as durable workflows. Deploys, crashes and provider outages pause them, nothing starts over, and a retry never pays twice for a value already made.
Failures that look like success
Empty answers with a normal finish, streams that stop early, schemas each provider handles its own way. All detected and retried before anyone sees them.
Built not to guess
Answers can only come from your option lists. Where code cannot decide, judge models check output against your rules, and an agent that is not sure answers “uncertain”.
The data graph
A product sits in several taxonomies at once. Rules inherit down the tree, attributes feed each other and run in order, a change re-runs what depends on it, and variants stay consistent.
Every value shows its work
Each value is stored with the reasoning, the inputs and the agent version behind it. Nothing is overwritten, so anyone can see why a value is what it is.
Cost per value, engineered
The cheapest model and prompt that hold quality for each task, photos sized to what each model can use, prompts built for caching, and translations planned per language.
Every model release is a migration project
Re-run a model on its own past calls and it does not fully agree with itself. Telling a real gain from that noise takes this, every time.
- RecordEvery production call is kept with its full input and output.
- ReplayCandidate models run on real past work, across attribute types.
- JudgeJudge models compare every disagreement, in both orders, so position bias cancels out.
- Roll outCustomers move once a candidate wins, and every value stays traceable to its version.
Who owns what
Your team owns the judgement. Cernel owns the machinery.
Your experts spend their time on the decisions only they can make, and your developers stay on your own roadmap.
Your team
- Product standards and business rules
- Brand voice and what good looks like
- Which attributes matter in each category
- Which values go live automatically and which get a review step
Cernel
- Models, prompts and what they cost to run
- Evaluation and re-testing at every model change
- Infrastructure, monitoring and upkeep
- Integrations that keep your systems in sync
Built in-house, owned by IT
- Ticket
- Backlog
- Sprint
- Release
Every new attribute, value or fix waits on the IT roadmap. To the people who use the tool every day, it always feels half finished.
Cernel, owned by your product team
- Describe it
- Test on real products
- Approve
IT connects your systems once. After that, a new attribute is a sentence your product team writes, tests and approves.
Decide
Build if it makes you different. Buy if it keeps you running.
Building is the right call for some teams. Tick what is true for yours. Nothing you tick is stored or sent anywhere.
Building can make sense when
Buying makes more sense when
Tick what is true for your team, and see which way it leans.
Teams that bought
Live on the platform while an in-house team would still be hiring.
Molly & My
Fashion
- Improvement in search conversion rate
- +97%
- Faster time-to-market
- 4x
Firtal Group
Health and beauty
- Faster time-to-market for new products
- +80%
- More organic traffic to product pages
- +20%
A new supplier sheet used to mean weeks, if not months, of work. Now we're talking days.
Figures from the published customer stories. Outcomes depend on how each team uses its data.
Before you decide
Questions to answer before you build.
Whichever way you go, these decide what it costs to keep running.
Who re-tests every attribute when the model changes?
New models arrive all the time, and a model can disagree with its own past answers. Built in-house, someone has to re-test every attribute in every category before anything switches.
With Cernel, we do it on real past work before any customer moves, and every value stays traceable to the version that made it.
Who forecasts the AI bill as the catalogue grows?
Every product, attribute and language is a call with its own price, and every re-run as rules and agents change adds to it.
With Cernel, model and token costs are ours to manage. We bring them down lever by lever: the right model and prompt per task, photos sized per model, caching and translations planned per language.
How long from “we need a new attribute” to live?
Built in-house, a new attribute is a ticket on the IT roadmap, and it waits its turn.
In Cernel, your product team describes it in plain language, tests it on real products and approves it. Your developers stay on your own roadmap.
Who owns accuracy, and how is it measured?
Accuracy needs rules, checks on every output and a record of why each value is what it is.
In Cernel, you choose which values go live automatically and which get a review step. Review steps check every value and only ask a person when something fails or needs a decision, and each value keeps its sources, reasoning and agent version.
What happens when the people who built it leave?
An in-house build lives in the heads of the few people who made it. When they leave, the knowledge goes with them.
With Cernel, nothing changes for you. A fix or improvement made for one customer reaches every customer when it ships.
Can we build on top of Cernel instead?
Yes. Many teams buy the machinery and build what makes them different on top. The public REST API lets your own systems push products in, start enrichment and collect approved content.