Nine additional reliable attributes across 4.71M+ products.
Phinest replaced an expensive, opaque deterministic solution with a flexible, client-owned pipeline that extracts core product fields and unlocks new attributes for downstream use.
items run through the extraction and validation pipeline
reliable product attributes unlocked using enriched source data
most new attributes support EQV calculations; others support understanding and potential matching
The old solution was expensive, brittle, and opaque.
An external provider used deterministic rules and pattern matching to extract product attributes. Results were unreliable, and every correction meant reporting the problem, waiting for rule changes, and receiving another output without visibility into the process.
What Phinest shipped
- A unified pipeline for product descriptions and images from multiple third-party providers
- Initial extraction of size, unit, product form, pack count, and a bundle boolean
- Nine additional reliable attributes unlocked using enriched Amazon product data
- A model-estimated consumer-serving value by product type
- Controlled-vocabulary validation and an application for human review
Why it mattered
- The flexible replacement reduced the cost of working with the previous expensive deterministic solution
- Client-owned code and configuration made the extraction process inspectable and faster to change
- Most additional attributes support downstream EQV calculations; others improve product understanding and may support matching
Four explicit stages, not another black box.
The Python pipeline uses Snowflake, Databricks, and Azure-hosted OpenAI endpoints. Its code and configuration are held by the client.
Inputs
Descriptions and images arrive in inconsistent formats from multiple third-party providers
Extraction
Text and vision models extract packaging attributes, detect bundles, and estimate a consumer-serving value by product type
Validation
Controlled vocabularies and bundled-product detection check the generated values
Control
The client owns the code and configuration, with a review interface and logged pipeline runs
Built for EQV calculations and richer product understanding.
Most additional attributes support downstream equivalized volume (EQV) calculations. Others add context for product understanding and may support matching.
Small-sample proof
The work began with a client data sample and iterative demonstrations before production investment
Reference comparison
Early results were compared with existing validated values and sometimes surfaced more specific information
Serving estimate
The model estimates a consumer-serving value by product type rather than treating it as extracted ground truth
Prove the hard part before funding the whole system.
Start with a representative sample, target attributes, and an agreed way to measure whether the extraction works.