In April 2023, Hyeonji Hwang founded Smoretalk to solve a common pain point for marketers. Their visual workflow for creating assets—which relies on references such as mood boards, past campaigns, and competitor ads—didn’t fit the text-prompt model of many AI tools.

The team’s first product was Flamel, an AI assistant that creates on-brand visuals from a reference image in minutes. As the team grew, they pinpointed a more specific challenge: generating ad banners with precise layouts and legible text.

This insight led them to develop a second, enterprise-focused product: the Ad Creative Agent. To improve their flagship tool, Flamel, while building this new product, CEO Hyeonji and her team realized they needed a smarter, more capable technical foundation.

The challenge: Scaling a product suite with technical precision

The Smoretalk team faced distinct challenges for each of their products. For Flamel, the team needed to refine its prompt tuning. “Transforming rough user prompts into optimized ones was often a two-step process requiring multiple tools,” Hyeonji explains. “If the user prompt wasn’t in English, we first used Google Translate. Then, we had to use a separate LLM to refine the translated text into an ideal prompt.”

Additionally, Hyeonji and her team also wanted to ensure Flamel users could make precise design edits without a steep learning curve. This meant refining Flamel’s overall ability to analyze reference images and generate on-brand assets, as well as improving an underperforming yet critical editing feature: object removal.

“We used a pipeline combining Stable Diffusion Inpainting with our own custom logic for object removal,” Hyeonji says. “But performance was poor and the output quality varied significantly.”

Developing the specialized Ad Creative Agent presented its own technical hurdles. “Many multimodal AI models still struggle to generate accurate and legible text in images,” Hyeonji explains. This was a major roadblock in creating a reliable tool that enterprises would trust. “We received ongoing improvement requests from clients who were testing the Ad Creative Agent,” Hyeonji shares.

The solution: Building smarter workflows with the Gemini API

“After experimenting with various models, Gemini stood out as a production-ready solution,” Hyeonji shares. “Gemini’s speed, cost-effectiveness, and reliability gave us a clear advantage over other third-party solutions.”

To launch improvements quickly and reliably, the team uses Google AI Studio to test new features before they go live. “We start by experimenting with new workflows in Google AI Studio,” Hyeonji explains. “Once we validate a rough process, we connect the Gemini API to our beta server, which mirrors our live product. If everything runs as expected, we then officially deploy the new workflows to our main server.”

This testing framework enabled the startup team to refine multiple workflows within Flamel, such as curating reproducible reference styles. “We use Gemini capabilities across this entire pipeline, from trend analysis to visual regeneration,” explains Hyeonji. The Smoretalk creative team first selects design trends and images. They then use Gemini’s multimodal capabilities to convert them into detailed text prompts that AI models can understand. This enables users to replicate a design style accurately without prompt engineering expertise.

Smoretalk's Gemini Workflow to Collect and Curate resources

Gemini also came in handy when refining Flamel’s prompt tuning. “Gemini streamlines the entire prompt tuning workflow into a single step, translating and optimizing prompts with greater accuracy and speed,” Hyeonji shares.

Hyeonji and her team also built smart editing workflows with Gemini’s vision-language model strengths so marketers without design skills can make precise edits. The team fixed Flamel’s underperforming object removal feature with the Imagen API, taking intuitive editing for non-designers one step further.

To improve their new Ad Creative Agent’s ability to handle layout, text generation, and product composition in banner ads, the Smoretalk team integrated the Gemini 2.5 Flash Image model into their pipeline.

“The Gemini 2.5 Flash Image model delivers outstanding performance in image editing tasks,” says Hyeonji. “It’s proven to be highly effective for composing and refining trendy banner drafts, particularly when combining product images or applying targeted edits. The fact that it was made available as an API right away meant we were able to quickly integrate it into our workflow and provide immediate, tangible value to our users.”

The results: Increasing user engagement and client trust

By building with Gemini, Hyeonji and her team were able to drive success across both of their products. Switching to the Imagen API increased monthly usage for Flamel’s object removal feature by 10X.

The team is also winning over enterprise clients with the improved Ad Creative Agent. “After integrating the Gemini 2.5 Flash Image model into our pipeline, we’ve reached a high level of client satisfaction and are now moving forward with discussions for full production-level contracts,” Hyeonji shares. “We’re also in discussions to scale this offering through partnerships with major brands.”

What’s next: Expanding capabilities and automating workflows

The Smoretalk team plans to expand Flamel’s capabilities into motion graphics and add new 2D illustration styles to their popular Figma plugin, Flamel 3D Icon Generator. “We believe this will allow Flamel’s powerful curation capabilities to be utilized across more domains, including education,” Hyeonji notes.

Her team will also use Gemini to improve Flamel’s reference-based image generation. “Flamel currently takes a single reference image. But we’re planning to use the Gemini 2.5 Flash Image model to enable more flexible creation by referencing multiple images simultaneously,” Hyeonji says.

For the Ad Creative Agent, the team’s focus is on acquiring more than 1,000 enterprise clients by 2026 while expanding its capabilities to short-form video. Hyeonji and her team are also integrating the Gemini API into internal admin tools to automate content curation. This, in turn, will help scale resources for both of their products faster.

Based on her experience, Hyeonji’s advice to other founders is to take advantage of production-ready AI solutions. “Gemini’s performance can deliver immediate, commercially viable value,” Hyeonji shares. “Combining Gemini capabilities with our understanding of user workflows is what allows us to provide the best possible experience and remain agile.”