Nothing moves until the data is right. alfa makes it right first.

Source recon

Before any build starts

Recon: what exists, what is missing, what it will cost

Problem. Every catalog project starts blind. Nobody knows how many products really exist, where the data lives, which platform and bot protection stand in the way, or how complete each field is.

Solution. A recon run detects the platform and protection, counts live products, checks structured data and pagination. With client files it joins every source on a verified key and scores coverage block by block: barcodes, colors, composition, descriptions, icons.

Impact. Three answers before any money is spent: volume, difficulty and fit to the target, with an effort estimate.

Build pipeline

Intake and build · one adapter per brand

The build pipeline behind every catalog

Problem. Every brand ships its data differently: live sites behind Akamai or Cloudflare, platform exports, Excel masters, archives. Building each catalog by hand does not scale and breaks on the next update.

Solution. A multi-process crawl layer with resumable runs, a hard timeout per page and a live control board. Each brand gets a thin adapter of four functions; everything else is shared core: normalize, map to the target template, build.

Impact. New brands become configuration, not projects. 14 watch brands, 4 sports brands and a jewelry catalog all run on the same core.

Reconciliation

Two sides, one verdict

A verdict for every item, before it goes live

• 4 verdicts: both, here, there, conflict
• 3-way source · master · live

Problem. The master, the supplier file and the live store never fully agree. Uploading without comparing means overwriting good data with bad.

Solution. Cross-match every item on an identity key across two or three sides, and give each one a verdict: in both, only here, only there, or conflicting.

Impact. The decision of what enters the live system is made before it enters, item by item, with an owner for every conflict.

Variant split

From model row to real SKU

Every real SKU, and only real SKUs

• 41,178 SKUs written in one catalog
• 0 invented variants

Problem. Suppliers send one row per model with axes like color, size, metal and finish. Stores need one row per real SKU.

Solution. Expand the axes, enrich, translate the parallel columns, and produce a gap report for everything that has no source.

Impact. One row per real SKU. If a SKU does not exist, the variant is dropped instead of invented.

Specs & filters

Feature, spec and filter layer

Specs, rating scales and smart filters

• 0 invented values

Problem. Shoppers filter and compare on specs that suppliers rarely provide in a usable form. Free text cannot be filtered.

Solution. Feature icons, spec mapping, rating meters (cushioning, flexibility, breathability, drop), tags, metafields and filters built on real attributes.

Impact. Product pages customers can filter and compare, not only read. Blank if there is no real source, never a guess.

Images, video & DAM

Autonomous DAM ingestion · multi-server

Every image on the right variant

• 26,275 images re-hosted, 0 broken

Problem. Images live in enterprise DAMs and media hubs (Adobe Dynamic Media, Bynder) with no clean link to the product, the color or the angle.

Solution. An autonomous harvester that reads the DAM, de-duplicates, and ties each image to its color and angle. The hero image of each color goes on every variant row, and video goes in as a gallery row.

Impact. Galleries that match the variant the shopper chose, with zero manual sorting. Two lessons are now rules: never de-duplicate by file-name family, and image sources are always full URLs.

Localization

Two layers · locked glossary

Native language, not a dictionary swap

• 2 layers: prose and specs
• 6 glossary priority levels

Problem. International brands landing in a new language. Generic translation flattens technical terms and reads like a manual.

Solution. Marketing prose goes through a language model per collection. Specs go through a deterministic pipeline with a locked glossary: keep-English, do-not-translate, whole values, tokens, terms and units.

Impact. Native pages in the target language, with brand, model, caliber, finish and certification names kept as they are. Every conflict is recorded.

Multi-agent review

Storefront sweep

A team of AI reviewers on every storefront

Problem. A live store with thousands of pages cannot be checked by eye, and errors hide in rendering, not in the file.

Solution. Parallel AI agents sweep the live storefront and report every rendering or data error with its source row.

Impact. One sweep recovered watches hidden by a single metafield error.

Validation gate

The central asset · before every write

Nothing reaches a live system without a PASS

• 143 checks before every write

Problem. One file that looks valid can silently wipe thousands of live fields: a blank in a merge import deletes a live value without an error.

Solution. A gate between build and delivery: locked rules, a certification suite by category, update-file checks, image-row checks and an image gate. Only changing columns go in an update file, and every blank is checked.

Impact. No file leaves without a PASS in every category, and a file that would blank a live field fails before it ships.

Delivery & routing

Each target in its own format

Store, POS, marketplace and ERP, each in its template

• 0 automatic publishes

Problem. Every target system wants its own format, and a wrong import is expensive to undo.

Solution. Approved files are routed to the store import, the POS template, the marketplace template or the ERP. Reports and boards are delivered as live web pages, public or password-gated.

Impact. Imports ready to run and reports the client can open. Editing and publishing are always two separate approvals.

Delta monitoring

After the catalog is live

Keep it right after launch

• 0 silent deletions

Problem. A catalog that was right on launch day drifts: new products, discontinued ones, price and spec changes upstream.

Solution. Find what changed since the build and scan only that. New pages are added, removed pages are marked discontinued and never deleted, stale records are flagged.

Impact. A safe update file every cycle, plus a report of what would have been deleted if the raw file had been uploaded. This is the monthly program.

Multi-source & marketplaces

Many sources in · many channels out

One catalog, many sources, every marketplace in its own template

Problem. Every marketplace has its own template, rules and required fields, and the data arrives from several sources that never agree.

Solution. A multi-agent scraper and PIM export pipeline reads every source, and one master is mapped to each marketplace template with its own rules.

Impact. One store mapped to four marketplaces, and product URLs turned into variant-level rows. Early-stage work, now part of the core.

CEO reports engine (beta)

Beta · working name

From raw ERP exports to a CEO dashboard, every number backed

Problem. Management reports are built by hand from ERP exports, and nobody can say where a number came from or how it was calculated.

Solution. Reads the ERP exports and produces a dashboard and an Excel file from one source. Every number shows its file, fields and calculation; a forecast is marked as a simulation. Read only, never writes to the ERP.

Impact. The manager gets a snapshot, sell-through, branch transfers, stock movements, what to order and exceptions in money, on top of the ERP and BI, never instead of them.