---
title: "How Botika runs full-stack generative AI on Modal"
description: "Botika builds agentic e-commerce teams that automate visual production for global fashion brands, using proprietary foundation models and a 100-terabyte image dataset. The company runs its entire full-stack AI operation on Modal, with the data pipeline processing 100 terabytes at less than 1% error rate across thousands of containers. Before Modal, the team struggled with Kubernetes fleet management and slow GCP Batch jobs; now a single researcher can launch 50 experiments daily using Claude, and they set up reinforcement learning infrastructure in less than a day. Modal handles autoscaling, cold starts, and multi-node training with RDMA, allowing Botika to double traffic in seconds, and they serve about 15 models on L4s, L40S, A100s, and H100s. The infrastructure is not Modal-specific, enabling the team to run open-source and custom frameworks without modification."
---

# How Botika runs full-stack generative AI on Modal

[Modal](https://yomu.fyi/company/modal) · Sep 2, 2026

**Type:** Problem & solution

## Summary

Botika builds agentic e-commerce teams that automate visual production for global fashion brands, using proprietary foundation models and a 100-terabyte image dataset. The company runs its entire full-stack AI operation on Modal, with the data pipeline processing 100 terabytes at less than 1% error rate across thousands of containers. Before Modal, the team struggled with Kubernetes fleet management and slow GCP Batch jobs; now a single researcher can launch 50 experiments daily using Claude, and they set up reinforcement learning infrastructure in less than a day. Modal handles autoscaling, cold starts, and multi-node training with RDMA, allowing Botika to double traffic in seconds, and they serve about 15 models on L4s, L40S, A100s, and H100s. The infrastructure is not Modal-specific, enabling the team to run open-source and custom frameworks without modification.

## Context

Botika builds agentic e-commerce teams that automate visual production for global fashion brands. The company needed a cloud infrastructure that could handle its full-stack AI operation, including a 100-terabyte data pipeline, training of proprietary foundation models, and production inference. Before Modal, the team struggled with Kubernetes fleet management and slow GCP Batch jobs, which hindered iteration speed and required significant DevOps effort.

## Approach / What changed

Botika adopted Modal as its sole compute platform, running data pipelines, model training, and production inference on Modal. The data pipeline processes 100 terabytes across thousands of concurrent containers with less than 1% error rate. Researchers launch experiments using Modal functions, and the team uses Modal’s multi-node training with RDMA for training foundation models. Inference uses Modal’s autoscaling to handle traffic spikes. The infrastructure is not Modal-specific, allowing the team to run open-source and custom frameworks without modification.

## Takeaways

- Modal’s autoscaling allows Botika to double or triple traffic in seconds without operational concern, eliminating the need for a separate queue like Redis.
- A single researcher can launch 50 experiments daily using Claude and Modal, compared to one or two before, enabling rapid iteration.
- Botika set up reinforcement learning infrastructure in less than a day using Modal primitives, with reward servers reused in production inference.

## FAQ

### How does Botika handle its 100-terabyte data pipeline on Modal?

Botika runs its data pipeline on Modal without a full orchestrator like Airflow. The pipeline processes 100 terabytes across thousands of concurrent containers with less than 1% error rate, using just application code files to express job dependencies.

### What infrastructure challenges did Botika face before Modal?

Botika struggled with Kubernetes fleet management, spending weeks chasing driver bugs, and found GCP Batch too slow for launching jobs. The alternatives required adopting a full orchestrator like Airflow or compromising on velocity.

### How does Botika train its foundation models on Modal?

Botika trains proprietary foundation models for 4K image generation using Modal's multi-node training with RDMA networking. They were early adopters of this feature and can run jobs across many GPUs for weeks. They also set up reinforcement learning infrastructure in less than a day using Modal primitives.

**Tags:** [AI](https://yomu.fyi/topic/ai), [Generative AI](https://yomu.fyi/topic/generative-ai), [GPU](https://yomu.fyi/topic/gpu), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Modal](https://yomu.fyi/topic/modal)

- Source: [Modal](https://modal.com/blog/botika-case-study)
- Source URL: https://modal.com/blog/botika-case-study
- Ingested by Yomu: 2026-09-20T07:47:18.426Z

[Read original post](https://modal.com/blog/botika-case-study)