Case study · Retail
AI-powered invoice processing for a global retailer
Cedar Retail Group
- Industry
- Retail
- Client
- Cedar Retail Group
- Stack
- Python, OpenAI, Anthropic, PostgreSQL, AWS, Temporal
Challenge
What we were solving.
Cedar's AP team was processing 18,000 invoices per month across 22 supplier formats, with a 6.4% error rate that triggered payment disputes and damaged supplier relationships. The team had grown to 27 people and was still missing the monthly close.
Solution
What we built.
We built an AI-powered invoice processing pipeline that ingests documents in any format, extracts line items with a fine-tuned model, validates against purchase orders, and surfaces exceptions to operators in a queue UI.
The system is built around an evaluation harness that catches model regressions before they reach production.
Results
What changed.
Invoice processing time reduced from 8 minutes to 28 seconds
Error rate dropped from 6.4% to 0.4%
AP team redeployed to higher-value work, with no layoffs
Model evaluation harness now catches regressions in CI
Technologies
Related projects
Northwind Health
Clinical decision support for a 14-hospital network
1.2M patients now have a unified clinical record across the network Clinician reconciliation time reduced from 40 minutes to 4 minutes per shift Risk score accuracy validated at 92…
Atlas Federal Credit Union
Migrating a regional bank to a multi-account AWS platform
CI/CD pipeline duration dropped from 90 minutes to 7 minutes On-call coverage expanded from 3 engineers to a 22-engineer rotation Cloud spend visibility improved to per-product lin…
Lumen Manufacturing
Predictive maintenance for a global manufacturing fleet
Unplanned downtime reduced by 38% in the first six months False alarm rate reduced from 71% to 6% Maintenance cost per line decreased by 22% ML pipeline scales to 240 production li…
Want a similar outcome?
Tell us about your system and we’ll share a relevant engagement model in a 30-minute call.