Descriptions:
Samir Mody, head of AI engineering at The Browser Company of New York, traces the company’s journey from Arc — a beloved but incremental browser — to Dia, an AI-native browser built from the ground up around LLM capabilities. The talk covers the organizational and technical challenges of shipping AI products at high iteration speed, with particular attention to the internal tooling that enables everyone from engineers to the CEO to prototype and refine features using their full personal browser context.
A centerpiece of the talk is JEBA, the company’s internal prompt optimization system. Rather than iterating along a single prompt-refinement path, JEBA uses a reflective prompt mutation and selection technique that explores a broader space of prompt variations simultaneously, optimizing against custom scoring mechanisms. Mody describes two explicit development phases: a high-breadth ideation phase where many ideas are rapidly dog-fooded internally, followed by a disciplined refinement phase involving eval collection, hill climbing through automated prompt optimization, and quality gates before shipping.
Mody also frames model behavior design as a formal craft discipline at the company — encompassing behavior design (tone, style, response shape), data collection for measurement and training, and model steering (prompt engineering, context window design, model selection). Dia currently features a memory knowledge graph, a computer use mechanism that went through tens of design iterations, and a personalized assistant that builds a user profile over time. The talk is a grounded, experience-driven account of what it actually takes to ship AI-native consumer products responsibly.
📺 Source: AI Engineer · Published December 19, 2025
🏷️ Format: Workflow Case Study
