Ishiros
Product Enrichment

Enrich Your Catalog
Without the Manual Work.

AI generates titles, descriptions, attributes, categories and units of measure from your raw PIM records bound by your brand rules and a human-in-the-loop quality gate.

01
[
]

What We Automate

Comprehensive enrichment from titles to SEO.

Scroll through process arrow_downward

01 //

Product Titles

Consistent naming conventions (brand → type → variant → size) in multiple languages, respecting channel limits.

02 //

Copywriting

Marketing descriptions for the card, technical descriptions for PDPs, and bulleted benefits in your tone of voice.

03 //

Attributes

Structured JSON: color, material, ingredients, dimensions strictly mapped to your attribute schema.

04 //

Categorization

Multi-level mapping to your category tree, plus tags for smart navigation and faceted filtering.

05 //

Units of Measure

ml vs g vs pcs automatic contextual detection, display normalization, and typo correction.

06 //

SEO Meta

Meta titles, descriptions, image alt text, and schema.org Product structured data for Google visibility.

CALCULATE YOUR SAVINGS

PIM industry baselines for manual enrichment land around 4 minutes per SKU. Plug in your catalog size and see how much time the pipeline gives back to your team.

products
100 50,000 100,000+
Manual enrichment
333 h
≈ 41.7 working days
With Ishiros pipeline
5.6 h
≈ 0.7 working days
Time you get back
328 h
≈ 41 working days
Pipeline speedup
~60×

Baseline: ~4 minutes per SKU for manual enrichment (titles, descriptions, attributes, categorisation, SEO meta) — consistent with figures published by Salsify, Akeneo, and Productsup. Pipeline benchmark: ~4 seconds per SKU on real customer catalogs. An 8-hour working day = 480 minutes.

HOW IT WORKS

Six steps. Deterministic where it has to be, AI where it pays off. Your hand-approved examples are the quality benchmark.

Step 1
01 · Ingest

Pull raw data

CSV from Akeneo, Excel from suppliers, JSON from REST APIs, photos from S3. Everything maps onto your unified schema.

Step 2
02 · Rules

Brand rules as code

Naming conventions, allowed categories, tone of voice, banned words, unit formats. Encoded not suggested to the model.

Step 3
03 · Generate

Few-shot on your examples

LLMs with the most similar examples from your hand-approved catalog. Different models for different fields.

Step 4
04 · Validate

Quality gate before publish

Length checks, unit checks, semantic agreement between title and description. Suspicious goes to review, the rest flows straight to PIM.

Step 5
05 · Sync

Push back to your system

Akeneo, Pimcore, Shopify, Magento, WooCommerce or a custom PIM via API. Incremental only changed rows.

Step 6
06 · Quality loop

Learn from corrections

Every human correction becomes a new few-shot example. The pipeline gets sharper every week without retraining.

INPUTS → OUTPUTS

Whatever shape your supplier data arrives in, we map it onto a clean, channel-ready record.

What you give us
  • Supplier Excel/CSV in any column order, with merged cells, multilingual values and inconsistent SKUs.
  • Product photos (S3, Drive, FTP) for visual attribute extraction — colour, material, packaging type.
  • Raw GTIN/EAN codes, manufacturer part numbers, supplier model names.
  • Existing PDF datasheets, technical specs and brochures.
  • Partial PIM exports — gaps and inconsistencies are part of the input, not a blocker.
  • Brand guidelines: tone of voice, banned words, naming patterns, category tree.
What we send back
  • Channel-ready titles with strict length limits per surface (Google Shopping, Amazon, your storefront, marketplaces).
  • Marketing copy for the listing card and a longer technical description for the product page.
  • Strictly-typed attribute JSON: colour from an enum, dimensions in mm, weight in g, material from a fixed list.
  • Category path mapped to your taxonomy plus secondary tags for search facets.
  • SEO bundle: meta title, meta description, image alt text and schema.org Product JSON-LD.
  • An audit trail per field: source, model used, confidence score and reviewer.

BUILT FOR THESE SCENARIOS

Different catalogs, different rules. The pipeline adapts — same architecture, your taxonomy.

Fashion & apparel

Size charts normalised across brands, fabric composition extracted from spec PDFs, season tags, fit descriptors picked from a controlled vocabulary so search facets stay clean.

Beauty & cosmetics

INCI ingredients parsed and standardised, allergen flags surfaced, volume normalised to ml, claims (vegan, cruelty-free) restricted to a verified list, descriptions per market language.

Electronics & tech

Specs extracted from manufacturer datasheets — connectors, voltages, throughput. Compatibility tables generated. Models grouped into families. Hard limits prevent hallucinated specs.

Grocery & FMCG

Net quantity vs. drained weight handled correctly, allergens called out per regulation, country-of-origin extracted, storage instructions standardised, halal/kosher flags carried from supplier data.

B2B & industrial parts

Manufacturer part numbers preserved exactly, cross-references to OEM equivalents generated, technical drawings linked, units kept in the engineering convention (Nm, bar, IP rating).

Marketplaces & aggregators

One source record, multiple channel-shaped exports. Each marketplace gets the title length, attribute schema and category code it expects — no manual reformatting per channel.

WHERE IT PLUGS IN

No rip-and-replace. The pipeline reads from your sources and writes back into the systems your team already uses.

Source

Where data comes from

Akeneo, Pimcore, custom MySQL/Postgres PIMs, supplier portals, Excel/CSV drops on SFTP, REST APIs, ERP exports (SAP, Oracle, Pantheon, Minimax), Google Drive folders and S3 buckets.

Engine

How the work happens

Multi-model orchestration: small models for normalisation, larger models for prose, vision models for photos. All wrapped in a deterministic rule layer with retries, validation and a human review queue.

Destination

Where data lands

Shopify, Magento, WooCommerce, BigCommerce, Salesforce Commerce Cloud, your custom storefront via REST/GraphQL, marketplace feeds (Google Shopping, Amazon, eMAG, Limundo) — or back into the same PIM you started in.

~4 min

per product, by hand

~8 s

per product, in pipeline

94%

first-pass acceptance

30+

supported languages

FREQUENTLY ASKED QUESTIONS

What if the AI generates an inaccurate description? expand_more
The validation gate intercepts anything that breaks the rules wrong unit, oversized title, an ingredient not in the input. Suspicious products go to a human review queue instead of straight to PIM. Every correction becomes a new training example.
Does it work across multiple scripts and writing systems? expand_more
Yes. Multilingual by design — generate in parallel across Serbian (Latin & Cyrillic), English, German, French, Italian, Spanish, Russian and 25+ other languages, including non-Latin scripts. Each language has its own examples and conventions.
How fast can we go live? expand_more
An MVP for one product category typically ships in 2–3 weeks, including schema analysis, brand-rule definition, and a 100-SKU pilot. Full rollout 4–8 weeks.
Does data leave the EU? expand_more
We default to leading LLMs, all on zero-retention contracts. For strict GDPR: open-source models (Llama 3.1, Mistral) on EU infrastructure Frankfurt or Belgrade.

Send 100 SKUs. We send them back enriched.