How forward-deployed data scientists built custom classification models and embedded them into daily catalog workflows for a national grocery cooperative.
Retail item descriptions are frequently abbreviated, inconsistently punctuated, duplicated across naming conventions, or associated with invalid product hierarchies. Examples include shortened terms like "chz" for "cheese" and conflicting department labels like "ALCOHOL AND REGULATED ITEMS" vs "ALCOHOL & REGULATED ITEMS."
Manual classification requires catalog specialists to interpret shorthand, identify the intended hierarchy, search for existing PLUs, and resolve conflicting master data. In the source dataset, 3,684 of 135,683 records had an invalid hierarchy and were excluded from model training—representing 2.7% of available samples.
The production objective: reduce repetitive catalog research while preserving human control over low-confidence or novel classifications. Planning assumptions targeted a 55% reduction in manual hierarchy-classification effort and a 35% reduction in PLU research time.
Forward-deployed data scientists embedded directly into the grocery cooperative's catalog operations. Not as external consultants delivering a model and leaving—as embedded engineers building custom solutions for the cooperative's exact data constraints.
The team built custom in-house language expansion models to resolve retail abbreviations. They developed custom in-house text encoders and a multi-output hierarchical classifier with separate prediction heads for department, sub-department, and category. They created custom class-weighting functions to handle sparse categories and ran automated in-house hyperparameter tuning.
For PLU mapping, they built custom in-house sentence-embedding models, an in-house similarity index for efficient cosine-similarity lookup, and a custom PLU bucket classifier that assigns valid ranges when no sufficiently similar item exists.
This wasn't a pre-trained model from a third party. It was forward-deployed data science: custom models, custom normalization logic, custom validation rules, and confidence-based routing—all built for one client's exact catalog structure.
AI Enablement Engineers made sure the models actually got used by catalog specialists. They built confidence-based routing: high-confidence, hierarchy-valid results auto-applied; medium-confidence predictions presented as recommendations; low-confidence or novel items assigned to manual review queues.
They designed human-in-the-loop workflows so catalog specialists could confirm, correct, or reject model predictions. They added normalization caching so expanded descriptions were stored with the model version, original text, normalized text, and timestamp—preserving lineage and avoiding repeated language model calls.
The result: catalog specialists can process normalized product descriptions, validate hierarchy combinations, review semantically similar PLUs, and confirm bucket assignments—all within their existing workflow. The system reduces repetitive research without removing human judgment.
Data Sources: Combined product tables (item descriptions, hierarchy values), pivot/master hierarchy tables (valid combinations), customer and member-retailer-facing descriptions (for normalization and semantic matching).
Core Components: Custom in-house language expansion model, custom in-house hierarchy classifier with shared text encoding and separate prediction heads, custom class-weighting for sparse categories, custom in-house sentence-embedding model, in-house similarity index, PLU bucket classifier with exact-match and cosine-similarity rules.
Hierarchy Coverage: 7 unique departments, 23 unique sub-departments, 265 unique categories. Sub-category excluded due to insufficient training volume.
Training Scale: 135,683 product records evaluated. Hierarchy classifier consumed approximately 18 GPU-hours across three training iterations. PLU bucket model consumed approximately 1 GPU-hour across two iterations.
Production Targets: Normalize and score 100,000 item descriptions in less than 45 minutes on a dedicated inference GPU, with at least 99% job completion and complete prediction logging.
The strongest results are at the broader hierarchy levels: 97% validation accuracy for department and 94% for sub-department. Category and end-to-end hierarchy accuracy require additional labeled data, constrained predictions, and evaluation through the full preprocessing pipeline.
Manual testing on unseen descriptions produced 85% department accuracy, 74% sub-department accuracy, and 63% category accuracy. When only fully valid department, sub-department, and category combinations were counted as correct, end-to-end accuracy was 60%.
Planning estimates assume a 55% reduction in manual hierarchy-classification effort and a 35% reduction in PLU lookup effort when deployed with confidence-based review.
Forward-deployed ownership: Data scientists didn't deliver a model and leave. They embedded in the catalog operation, owned the normalization logic, managed the hierarchy validation, monitored the similarity index, and iterated based on catalog specialist feedback.
AI enablement for adoption: AI Enablement Engineers optimized the system for how catalog specialists actually work. They refined confidence thresholds, improved review queues, added normalization caching, and ensured predictions matched catalog governance requirements.
Outcomes-based iteration: The team measured department accuracy, sub-department accuracy, category accuracy, PLU bucket accuracy, and hierarchy-validity rate. They iterated on the components that moved those numbers.
Data quality first: The team rejected 2.7% of records with invalid hierarchies before training. They standardized conflicting labels, validated hierarchy combinations, and built custom class-weighting for sparse categories. High validation accuracy doesn't remove data-quality risk—it requires it.
Want to see how forward-deployed data scientists and custom AI models can improve your catalog operations?
SCHEDULE CONSULTATION