Botika Leverages Modal for Scalable Full-Stack Generative AI in E-Commerce
- Gate
- Machine Learning
- Departs
- Crew
- William Rodriguez
- Passengers
- 350
Botika's innovative approach automates the visual production pipeline for fashion brands, utilizing a custom-built full-stack AI on Modal for enhanced efficiency.
Botika serves as a pioneer in automating visual production for global fashion brands by creating agentic e-commerce teams that proficiently handle the entire process from image generation to personalized content delivery. At the heart of this operation lies a full-stack artificial intelligence (AI) system, featuring proprietary foundation models with tens of billions of parameters along with a robust data pipeline that processes an astonishing 100-terabyte dataset. This intricate system is effectively managed on the Modal platform, which plays a crucial role in Botika's operations.
“I don't think Botika would have existed in its current form without Modal," says Eran Dagan, Botika's co-founder and CEO. "We would have needed probably twice as many people.”
Transition to Advanced Cloud Solutions for AI Workloads
The journey of Botika began in 2018 when its founders ventured into generative AI well before suitable tools were available. They undertook the challenge of building their own tooling for workloads, including node autoscaling, cold start mitigation, and GPU fleet management. In 2023, Dagan sought an advanced cloud solution, evaluating existing GPU cloud platforms only to find them inadequate for their needs. Modal emerged as a solution when Dagan discovered it on an AI infrastructure chart.
“I remember the moment I typed modal run and it just worked. I was like, OK, this just saved me a week of work right here and now.”
With Modal, Botika streamlined its operations from data management to operational inference.
Efficient Data Pipeline Management Without an Orchestrator
Botika’s data pipeline transforms raw fashion images into structured training data through a dozen steps: feature generation, filtering, tagging, clustering, and aggregation. The pipeline incorporates various AI models, including both open-source Vision Language Models (VLMs) for tagging and specialized classifiers developed in-house. Previously executed on GCP Batch, the transition to Modal significantly shortened the processing time, allowing the team to complete tasks far quicker than before.
“By the time a job boots on GCP Batch, you could have launched a thousand jobs on Modal," Dagan says. "The logic of 'once this job is done, run this one' is so easy to express on Modal, and so difficult without a full-blown orchestrator anywhere else.”
Thanks to Modal, Botika can manage a 100-terabyte dataset with a less than 1% error rate, all while utilizing the diverse range of GPUs Modal offers, essentially processing thousands of concurrent jobs.
Accelerated Research and Development
The research component of Botika’s model training has also flourished on Modal, enabling researchers to conduct thousands of experimental training runs with unprecedented speed. Each researcher now easily runs about 50 experiments a day, compared to one or two previously. This breakthrough is primarily due to the ability to integrate rapid iteration into their workflow.
“You give a smart person, and an AI agent that translates that person's thoughts, access to very fast and very responsive compute, and you just get a ton more work done," Dagan explains. “Without Modal, the research team would be six months in the past.”
Botika's capability to train its foundation models for high-resolution image generation on Modal further underscores their pioneering approach, allowing jobs that typically take weeks to be completed more efficiently without requiring the team to set up and maintain extensive infrastructure. The team was among the first to use Modal's multi-node training feature, gaining access to advanced networking capabilities from the outset.
Dynamic Inference Capabilities
In production, Botika operates around 15 models across various GPU types, including L4s, L40s, A100s, and H100s. Modal’s inherent features like queuing, autoscaling, and solving cold start issues replace traditional dependency on Redis or similar queuing systems. This ability to scale elastically at a moment’s notice is particularly advantageous, as AI traffic patterns often require rapid adjustments to demand.
“We can double or 3x traffic in seconds and nobody knows about it. Nobody at the company cares from a technical standpoint, we know Modal will just work.”
An Integrated Cloud Environment
Botika's reliance on Modal reflects a departure from traditional cloud solutions. As Eran Dagan notes, the functionality provided by Modal covers nearly all aspects of their operations. This cohesive approach enables applications to be deployed without the need for a specific Modal configuration, allowing the team to utilize both open-source and proprietary frameworks without encountering barriers. This flexibility enhances productivity since tasks can be managed efficiently through concise commands.
Every engineer joining the team quickly recognizes the advantages of this integrated system, noting the stark difference in productivity compared to traditional methods like Kubernetes.
Ultimately, Botika’s experience emphasizes the growing significance of specialized platforms like Modal that cater to the unique demands of AI-driven businesses, enabling teams to achieve operational excellence without the complexity that often weighs down innovative processes.
