Mihail Eric says his first startup idea around data quality and labeling was too late and too crowded, even after meeting with many VCs.
2
Storia AI moved from an AI storyboarding application toward developer infrastructure for image generation, with APIs and tools for customization and editing.
3
Mihail Eric sees LLMOps as mostly MLOps applied to larger models, with data quality, serving, evaluation, and infrastructure concerns still largely familiar.
Summary
Mihail Eric talks with Demetrios Brinkmann about his path from Amazon Alexa and Stanford NLP to founding startups. His first attempt focused on data quality and labeling, but he found the market crowded and struggled to raise funding. He then worked with his co-founder Julia to build Rick and Mortify, a multimodal Rick and Morty episode generator. Interest in that project led them toward AI-powered storyboarding and eventually Storia AI. Mihail argues that image generation still lacks developer-friendly infrastructure. Off-the-shelf models are not enough for many products, which need fine-tuning, LoRAs, editing, and other forms of adaptation. He also discusses AI hype, the effect on Hollywood, and why human sentiment and experience still matter in creative work. On MLOps and LLMOps, he says the models and infrastructure can be different, but many underlying concerns remain the same, especially data.
Mihail's first startup entered data quality and labeling too late
Mihail Eric describes his earlier company as an MLOps and data-focused idea built around structured data, labeling, and data quality. He says the timing was poor because too many companies were already in the space, and the data was not differentiated enough. He still believes the underlying idea has merit. Improvements should run through the full loop, from annotation to training to serving and model predictions, with information from predictions feeding back into labeling. He says companies often begin with one narrow part of the stack and then expand in either direction to own more of the value chain.
A multimodal prototype led Mihail and his co-founder to Storia AI
After meeting his co-founder Julia, Mihail built a small project that combined text and image generation. Rick and Mortify let users choose Rick and Morty characters, provide a premise, and generate a comic-like episode with dialogue and visuals. The project attracted tens of thousands of users in a short period. That distribution brought a conversation with a director and screenwriter, who suggested storyboarding as a useful application. Mihail and Julia followed that opportunity and built an AI-powered storyboarding tool, later developing it into the products they work on at Storia AI.
Image generation products need more than a model API
Mihail says a company cannot build a lasting product by taking a vanilla Stable Diffusion model off the shelf. Applications often need fine-tuning, many LoRAs, and a way to adapt the output to a particular style or industry. Generated images also have a last-mile problem. Customers may need to fix or modify an image before they can use it. He sees the surrounding ecosystem as underdeveloped, especially for developers who need production-ready building blocks rather than only artist-focused interfaces.
Storia AI is building focused APIs and example applications
Mihail explains that Storia AI began with a focused set of image-editing APIs instead of trying to support every possible feature. The company is also building applications that show what those APIs can enable, such as a browser editor and a Discord bot that can work alongside Midjourney. He wants developers to use the same building blocks for many products, including print-on-demand services, video game editors, and other tools. The company plans to add video capabilities gradually while keeping the developer experience consistent.
Unexpected vertical use cases come from customer conversations
Some of Storia AI's use cases were not part of Mihail's original plan. One customer wanted to let users create patterns and turn them into custom wrapping paper. Other examples included adding watermarks and helping with visual concepts for housing projects. Mihail says these conversations pull the company into applications it would not have imagined on its own. This is why he describes the infrastructure as a more horizontal layer, even though certain areas such as video games, entertainment, architecture, interior design, and marketing show especially clear demand.
AI hype creates unrealistic expectations for technical progress
Mihail has watched several AI hype cycles since entering the field more than a decade ago, but he says the last one or two years have been unlike anything he has seen. Consumer applications such as ChatGPT and DALL-E made the technology visible to people outside AI. He expects the hype to contract as people move from believing everything is possible to asking where the technology actually works well. He worries that confident but shallow commentary can raise expectations so high that people become disappointed when difficult research and economic problems take years to resolve.
Creative AI should assist people while preserving human experience
Discussing Hollywood, Mihail says workers have valid reasons to be concerned about rapid change and about compensation, credit, and attribution. He also argues that people in the arts need to believe there is something fundamentally human in creative work, such as human sentiment, memory, and lived experience. He expects AI tools to enter smaller parts of the production process first and make people more productive. He does not think current short video clips indicate that visual-effects teams can soon be removed from film production.
Mihail agrees with the idea of replacing the small 'ML code' box in D. Sculley's technical-debt diagram with 'LLM code'. Larger models create new infrastructure problems, such as hosting models with billions of parameters, but many other concerns remain familiar. Data is still central to improving systems, including better data, captions, and training material. Mihail points to companies that moved into LLM work without rebuilding their businesses from scratch. In his view, LLMOps is largely MLOps supporting a new class of models, with some changed infrastructure requirements and much more customer interest.
"You have to kind of start from like labeling, trying to solve the labeling problem, and then expect that once you solve labeling and you build out through the rest of the stack you'll start to hit into issues around increasing the quality of the data through a training run through serving."Mihail Eric09:56
Who should watch
You are building a product on top of image or video generation models and need to understand where model APIs stop being enough.
You are weighing a startup idea in MLOps, data quality, or generative AI and want a candid account of timing, differentiation, and distribution.
You work on ML infrastructure and want Mihail Eric's view on whether LLMOps is a genuinely new discipline or a new set of models inside familiar MLOps concerns.
