Descriptions:
Manos Koukoumidis and Stefan Webb introduce VibeML and its UMIE agent platform at AI Dev SF 2026, making the case that enterprises are rapidly shifting from renting generic intelligence via APIs (OpenAI, Anthropic, Google Gemini) to owning specialized models tuned for their specific use cases. Koukoumidis, who previously worked on Gemini at Google, argues that specialized models can be 10 to 100 times smaller and more cost-efficient than general-purpose frontier models while delivering higher quality on targeted tasks—citing Kiro and Intercom as recent public examples of companies building coding-specific models that outperform GPT-4 class models at a fraction of the operational cost.
The UMIE platform acts as an AI-driven model factory: a user describes a task, and a conversational agent guides them through defining evaluation metrics, synthesizing training data, selecting a baseline model, running fine-tuning, and measuring improvements—all without requiring deep ML expertise. The live demo shows the agent proposing evaluation criteria (completeness, conciseness, format adherence, and faithfulness) for a news summarization task, creating LLM judges from scratch, synthesizing stratified test data, and fine-tuning Qwen 3.5 4B against those judges.
The session is particularly relevant for enterprise ML teams frustrated by the time and expertise required to build custom models. VibeML positions itself as a model factory that compresses a process that previously took months into hours, and lets any engineer in an organization build specialized, privately owned models without deep ML backgrounds.
📺 Source: DeepLearningAI · Published May 22, 2026
🏷️ Format: Keynote Launch
