The company is a leading food delivery and logistics platform that connects 20 million customers with local restaurants, grocery stores, and retailers, operating across 40+ countries. The platform handles diverse product categories requiring precise tax classification at scale.
Challenge
Precise tax categorization of items being sold on the platform is critical to ensure legal compliance and protect profit margins from costly government audits. With a growing number of items, the company faced a critical challenge in scaling its tax categorization system across restaurant and non-restaurant items. The existing process, largely dependent on manual tools like Google Sheets, achieved only ~55% accuracy, leading to potential financial risk, inconsistent tax application, and poor merchant and customer experience.
Additionally, there was no structured workflow for labeling and review, no auditability, and no visibility into team performance or operational efficiency. The company planned to expand labeling from predicting 60 tax categories for restaurant items to 600+ tax categories for non-restaurant items.
The company resorted to a third-party dataset labeling platform already available on the market to overcome the limitations of Google Sheets. However, the system was slow, click-heavy, and did not support bulk operations, making large-scale work inefficient. Hence, the tax labeling process was unreliable; workflows were not easily configurable, and it was not extendable to support the company’s future requirements, such as complex tax labeling workflows or ML model integrations.
Solution
The auto-labeling tool standardizes the data annotation process through intuitive labeling and review interfaces enriched with metadata and taxonomy-based navigation. Users could efficiently label, review, approve, comment on, or escalate items, while administrators gained the flexibility to allocate and rebalance workloads as priorities shifted.
Real-time dashboards provided end-to-end visibility into annotation progress, enabling better operational control and faster decision-making. The tool also leveraged LLM-powered auto-review capabilities to automatically process high-confidence items, significantly reducing manual review effort and accelerating turnaround times.
With built-in audit trails, continuous feedback mechanisms to improve ML model performance, and user productivity and quality tracking, it transformed a fragmented, labor-intensive workflow into a scalable, data-driven operation that improved efficiency, consistency, and governance.
The different technologies used for this engagement included,
React JS
Flask, Python
LLM Integrations
ML-based classification Models
PostgreSQL (DB)
Docker, Amazon EKS
Okta (RBAC)
Impact Delivered
The implementation of the auto-labeling tool streamlined and accelerated the tax-labeling process while improving visibility and control. Operational efficiency improved by reducing the labeling time by 50%. There was stronger quality control and a clear path toward achieving 90–95% tax categorization accuracy at scale while laying out the foundation for AI-assisted automation in the company’s tax infrastructure.
