Botika runs a 100-terabyte fashion AI stack on Modal, minus Kubernetes

Why it matters

Botika shows how venture-backed AI researchers can buy away cloud operations and direct scarce engineering time toward proprietary models, product quality and distribution.

Botika co-founder and CEO Eran Dagan has moved nearly every compute-heavy part of the fashion AI developer onto Modal, betting that rented infrastructure can give a relatively small research organization the operating range of a much larger lab.

A Modal customer case study published September 2nd details the extent of that bet. According to Modal, Botika uses its platform to process a 100-terabyte image dataset, train proprietary foundation models with tens of billions of parameters and serve roughly 15 models in production. Those workloads span L4, L40S, A100 and H100 GPUs.

Dagan's reported payoff is organizational. He told Modal that Botika would probably need twice as many people to operate its current stack without the infrastructure platform. That estimate comes from the customer featured in a vendor's sales document, rather than an independent staffing analysis. It still captures the choice Dagan has made: Botika directs engineering time toward fashion models and product features while Modal handles queues, autoscaling, cold starts and GPU capacity.

Dagan and co-founder Yarin Didi Meir began working in generative AI in 2018, according to Modal. Startup Nation Central lists Botika as founded by the pair in June 2019. Botika's biography says Dagan served in Israel's Unit 8200, built his first startup at 16 and scaled a global television community by 18.

That early start meant the founders encountered the infrastructure problem before managed AI platforms had matured. Botika built node autoscaling, cold-start mitigation, Docker image optimization and GPU fleet management internally. Dagan recalled losing weeks to unhealthy Kubernetes nodes and driver bugs as the workloads grew.

The founder's infrastructure bet

In 2023, Dagan began testing alternatives to Botika's existing setup. Botika's data pipeline was running on GCP Batch, where processing a large dataset involved launching virtual machines, running scripts and adjusting scaling parameters manually. Adopting an orchestrator such as Airflow would have added another system for Botika's engineers to operate.

Dagan found Modal on an AI infrastructure market chart and requested beta access. "I remember the moment I typed modal run and it just worked," he told Modal.

That simple developer experience became the basis for a broad migration. Botika now uses Modal for data preparation, model experimentation, multi-node training, reinforcement-learning infrastructure and production inference. Dagan estimates that compute represents 99% of Botika's business and says that 99% of its compute runs on Modal.

Botika's data pipeline turns raw fashion imagery into usable training material through feature generation, filtering, tagging, clustering and aggregation. Modal says the process calls roughly a dozen models, including open-source vision-language models for zero-shot tagging and classifiers developed by Botika. Each workload can run in a separate environment and on a GPU selected for that task.

Modal reports that Botika processes the 100-terabyte dataset with an error rate below 1% across thousands of concurrent containers. The performance figure has not been independently benchmarked, and the case study does not define which failures count toward the error rate. It does show why Dagan wanted to remove infrastructure configuration from the research loop: every additional model and data-processing stage multiplies the operational work around the underlying research.

Fifty experiments instead of two

Modal says Botika researchers previously managed one or two experiments per person each day. A researcher can now launch as many as 50 short experiments, select promising results and scale the winners into larger training runs. Botika also connects Anthropic's Claude to the workflow so agents can help generate and execute experiments.

Botika uses multi-node jobs that run for weeks to train foundation models for 4K image generation. Botika also assembled reinforcement-learning infrastructure from Modal functions in less than a day, according to the case study. Reward servers used during training can be reused in the production inference stack, keeping data preparation, research and serving inside the same operating model.

Those throughput claims measure iteration speed rather than research quality. Fifty experiments are valuable when researchers can identify which changes caused an improvement and reproduce the result. They can also produce fifty inexpensive dead ends. Dagan's bet is that Botika's researchers, assisted by agents, can make that selection faster than an infrastructure-heavy organization can provision and monitor each job.

The approach extends into production. Modal says Botika serves about 15 models at a time and can double or triple traffic within seconds. Botika said in December 2025 that it had passed 3,000 active paying customers and had rebuilt its model infrastructure for faster generation and more consistent output. Both the customer count and scaling performance are Botika's own figures.

Venture money goes toward the models

Botika has announced $18 million in financing across two rounds. On January 16th, 2025, Botika disclosed an $8 million seed round co-led by Stardom Ventures and Secret Chord Ventures, with Seedcamp participating. Botika later said it raised a $10 million Series A led by Infinity Ventures in 2025.

The infrastructure decision determines what that capital buys. Maintaining a Kubernetes fleet, debugging drivers and building internal orchestration would consume engineers without producing a new image, video or customer workflow. Modal takes that work and gains a deeply committed customer running data, training and inference on one platform. Botika accepts the corresponding concentration risk because Dagan values research speed and a smaller operating staff.

Botika's product ambitions make the tradeoff understandable. Botika says it has expanded beyond generated on-model photographs into flat-lay conversion, footwear imagery, video generation and personalized e-commerce content. Each format introduces additional models, data processing and inference patterns. A fragmented cloud stack would force Botika's engineers to connect those pieces before researchers could improve them.

Modal's case study does not establish whether the arrangement produces the lowest raw GPU bill. It establishes Dagan's preferred use of scarce engineering time. Botika is building a vertically integrated fashion AI product, including proprietary foundation models, without also trying to become its own cloud provider.

For Dagan, the infrastructure layer has become almost deliberately unremarkable. An engineer writes a function, runs modal run and moves back to the model. After years spent chasing Kubernetes nodes, that lack of drama is the product he was buying.