Product-content intelligence at marketplace scale.

The challenge

India's largest marketplaces run catalogues in the tens of millions of SKUs. Every one of those products needs content — and not just a description. It needs content that works in multiple Indian languages, and content that has to satisfy four masters at the same time.

A human content team cannot hold quality, speed and cost together at that scale. The math doesn't work. What's needed is the output of a content operation delivered at the throughput of a pipeline — without losing the judgment that makes content good.

The four jobs of a product attribute

Most teams write a product description and move on. We map, for every product, the full set of attributes that make it work — anywhere from a few dozen to 80+ for an attribute-dense category like fashion: fit, fabric, occasion, pattern, neckline, length, and on and on. Each of those attributes is doing four jobs at once.

1. Search filtering. Faceted navigation — the shopper filters by size, fabric or spec and lands on the right product fast.
2. SEO. Richer pages, long-tail keyword coverage, and structured schema that search engines rank.
3. Customer decision. The right attributes let a shopper find and confidently choose, which lifts conversion.
4. AI discovery. Well-structured products are what LLMs and shopping agents retrieve and recommend. Most catalogues are invisible to AI commerce because their content was written for humans skimming, not machines reasoning.

That fourth job is where the market is heading, and where almost no one is ready. We don't write descriptions. We build the attribute intelligence that makes a product findable by a filter, a search engine, a shopper, and an AI.

The six-stage pipeline

1. Ingest. Marketplace exports, PIM and API feeds are normalized to a canonical product schema, with the attribute model configured per category — a saree, a smartphone and a packet of atta are described by completely different things.
2. Extract. An intelligence layer pulls structured attributes out of messy source data — titles, seller copy, specs and images.
3. Generate. Missing attributes and copy are generated from that structure, so every product reaches the full attribute set for its category.
4. Localise. Content is produced in each target language from the structured product data, not translated sentence by sentence from English.
5. Route. Confidence scoring auto-passes what the system is sure of and sends only low-confidence items to human reviewers.
6. Measure. Throughput, cost per product, review rate and listing quality are tracked continuously, so the pipeline is tuned on evidence.

The specifics of our models, prompts and pipeline internals are proprietary and stay that way. The principle is what matters: humans move from producing content to governing it, and the ceiling on throughput shifts from headcount to compute.

Why we generate from structure, not from English

The usual approach to multilingual catalogues is to write in English and translate. Every translation step carries the source's gaps and errors into the next language, and phrasing that reads well in English often reads stiffly once translated.

Our multilingual engine works the other way round. Because every product is first captured as structured attributes, each language is generated from that structure directly — with that language's own vocabulary, search terms and conventions. Quality doesn't degrade language by language, and adding a new language means configuring an output, not commissioning another translation pass over the whole catalogue.

Proven at the top, built to travel down

We proved this engine at the hardest scale in Indian commerce — the catalogues of Flipkart and Myntra, where the volume, the category breadth and the language requirements leave no room for a fragile system.

The same engine is built to travel. The economics of a fashion marketplace are nothing like those of a value grocery retailer or a smaller regional platform — different categories, different attribute sets, different turnover. The system adapts to all of them, because the attribute model is configured per category rather than hard-coded for one vertical.

What this means for your business

If you're a language or localization business, this is the same engine running on your translation memory and your linguists — an AI production layer on top of the moat you've already built. See how it works for LSPs →

If you're a content business, this is a discoverability capability you can productize for your own clients — the thing that makes their catalogues findable not just by shoppers, but by the AI that increasingly does the shopping.

Frequently asked questions