Written by Rutuja Lohar, Chandra Shekar, Ravi Mevcha, Gaurav Kankaria

Introduction

BigBasket, India’s largest online grocery platform, serves millions of customers daily across 40+ cities. With ~600 Fulfillment Centers (FCs) and ~150 Distribution Centers (DCs) managing 26,000+ SKUs, its supply chain is among the most complex in the country.

Ensuring the right products are available at the right time is critical. Even small forecasting errors can cause stock-outs (lost sales and unhappy customers) or overstocking (waste and increased costs).

To address these challenges, BigBasket partnered with Oneture Technologies to jointly modernize its demand forecasting system. Together, the teams designed and implemented a scalable, AI-driven forecasting pipeline powered by AWS services. Using Amazon SageMaker, Amazon S3, Amazon EC2, Amazon CloudWatch, and Amazon Redshift, the new solution improved accuracy, reduced training time by 85%, minimized inference delays, and cut infrastructure costs — building a future-ready supply chain backbone.

Key Challenges Faced by BigBasket

While BigBasket had begun using DeepAR (a deep learning–based time series forecasting algorithm) at the FC level, extending it to DCs revealed major hurdles:

• Accuracy Gaps – Statistical models at DCs achieved only ~50% accuracy versus the 98% business target.
• Long Training Times – Training one DC with 26K SKUs could take 3+ days.
• Inference Bottlenecks – Generating validation forecasts serially took 3+ hours for a single FC.
• Scalability Issues – Monthly retraining and forecasting across thousands of SKU-DC/FC combinations overwhelmed the legacy system.

In short, BigBasket needed a faster, more accurate, and scalable solution.

Joint Solution: Building a Modern Forecasting Pipeline on AWS

The journey to modernize BigBasket’s demand forecasting pipeline began as a true collaboration between BigBasket and Oneture. The teams came together with a common goal: build a forecasting system that was not only accurate, but also scalable and cost-efficient — all while keeping pace with the company’s rapid growth.

AWS Technical architecture diagram for demand forecasting solution

Building Scalable ML Pipelines with SageMaker

The first step was to rethink how models were trained and deployed. BigBasket already had experience with DeepAR, but scaling it to Distribution Centers posed hurdles. Together, the teams turned to Amazon SageMaker, which provided a managed, modular environment for building end-to-end machine learning pipelines. With SageMaker Pipelines, the teams designed reusable, parameterized workflows that could run simultaneously across multiple business units. This transformed what used to be siloed, serial processes into a streamlined system that could scale to dozens of sites in parallel.

Harnessing Elastic Compute (EC2) for Faster Training

But designing pipelines alone wasn’t enough. Training thousands of SKUs across Fulfillment and Distribution Centers required heavy computation. By leveraging Amazon EC2’s m5 family of instances through Amazon SageMaker, the teams dynamically adjusted resources depending on workload size — m5.4xlarge for smaller FCs and m5.12xlarge for larger DCs. This flexibility cut training time from three days to just 15 hours, making retraining cycles significantly faster. AWS’s elasticity meant the infrastructure scaled up when workloads were heavy and scaled back down when idle, ensuring cost efficiency.

Ensuring Consistency with Dockerized Environments

Consistency across environments was another challenge. To solve this, the teams built custom Docker images and deployed them within SageMaker. This guaranteed reproducibility, portability, and ensured that every experiment, no matter the size or site, ran smoothly without dependency mismatches.

Accelerating Validation and Inference with S3 and SageMaker

Once training was optimized, the teams tackled bottlenecks in validation and inference. Earlier, generating forecasts for 60 days ran serially and took hours. Using parallelized workers within SageMaker, along with efficient data storage and retrieval from Amazon S3, the process became far more efficient. Validation that once took three hours was reduced to 24 minutes, and test forecasts dropped from 92 minutes to just 19.

Raising Forecast Accuracy with Smart Evaluation

Accuracy was equally critical. Together, BigBasket and Oneture built a custom evaluation framework within Amazon SageMaker to track root mean squared error, quantile accuracy, and SKU-level metrics. With AWS enabling fine-grained experimentation and monitoring, the team introduced day-wise quantile adjustments that intelligently increased safety stock for SKUs prone to stock-outs. This innovation pushed availability to 98%, while keeping dead stock minimal.

Embedding Monitoring and Cost Optimization

Finally, to ensure the system could grow with BigBasket’s needs, the teams embedded monitoring and cost controls into the architecture. Amazon CloudWatch tracked logs and performance, automatically shutting down idle resources. By running workloads on EC2 Spot Instances, the teams achieved 30–40% infrastructure cost savings without compromising reliability. Amazon Redshift served as the data backbone, providing SKU- and FC-level transaction data that fed into the ML pipeline.

What emerged from this journey was not just a technical upgrade, but a co-created solution that married BigBasket’s deep supply chain expertise with Oneture’s cloud and AI experience, powered every step of the way by AWS. The result: a forecasting system that is fast, accurate, scalable, and future-ready.

Business Impact

The new DC-level forecasting pipeline co-created by BigBasket and Oneture on AWS transformed supply chain operations:

• 45% faster training: From ~60 minutes to ~25 minutes at DC level.
• 88% forecast accuracy: Achieved reliable stock availability with reduced dead stock.
• 30–40% lower costs: Through spot instances and elastic resource scaling.
• Future-ready backbone: Modular, automated, and scalable architecture that can extend across hundreds of DCs and FCs.

This collaborative effort ensured customers consistently found what they needed, strengthened supply chain reliability, and positioned BigBasket for continued growth.

Conclusion

By modernizing its demand forecasting system on AWS with DeepAR in SageMaker, BigBasket successfully solved challenges of accuracy, speed, scalability, and cost-efficiency.

The partnership between BigBasket, Oneture Technologies, and AWS delivered a solution that is faster, smarter, and scalable—helping BigBasket serve millions of customers more reliably every day.

This case study underscores how AI-driven forecasting on AWS can revolutionize large-scale supply chains, setting a new benchmark for retail in India and beyond

About the Authors:

Rutuja Lohar is a Machine Learning Engineer at Oneture Technologies with around 3 years of experience across the Data and Machine Learning landscape. She has a strong command of statistics, machine learning techniques, and ML models, and consistently applies her expertise to solve real-world problems.

Chandra Shekar is an AI/ML Engineer at Oneture Technologies, specializing in machine learning, Agentic AI, and AWS cloud. He has delivered solutions across demand forecasting, multi-agent systems, RAG applications, and FMCG analytics. He has consistently worked on turning complex AI into practical business outcomes.

Ravi Mevcha is a seasoned Digital Product Engineering Lead at Oneture Technologies, driving ideas from concept to launch. He has delivered multiple multi-million-dollar B2B and B2C platforms across Automotive, eCommerce, FinTech, and Procurement. He holds a B.Tech in IT from Gujarat University and an MBA from IIIT, Pune.

Gaurav Kankaria is the Chief Client Officer at Oneture Technologies and a Global Technical AWS Ambassador, recognized for his Data and AI expertise. A 6x AWS certified professional, he specializes in delivering AI-powered solutions for the Retail industry. Gaurav holds a B.Tech in ECE from VIT University and an MBA from ISB.