Julius AI founder on CBT-style prompting, slide generation, and why talking to AI beats typing
Key Points
- Julius AI launched slides last month and routes generation requests between image models for single graphics and code-based HTML/JavaScript for multi-slide decks, where style consistency matters at scale.
- Founder Rahul Sonwalkar is building Julius toward an ambient analyst model where users email tasks directly and receive finished presentations without switching tools.
- Image generation now handles presentation graphics reliably enough for text rendering, but code-generated slides remain superior for maintaining brand consistency across multi-page decks.
Summary
Julius AI is an AI-powered data analysis platform that lets users bring messy data and produce insights, reports, dashboards, and slides through natural language. Rahul Sonwalkar, the company's founder, describes the near-term roadmap as closing the loop between analysis and action: users will soon be able to email Julius directly to assign data tasks, positioning it as an ambient analyst for finance, operations, and marketing teams.
“My May 2026 approach to prompting is I've completely given up on like yelling at AI. I do more like cognitive behavioral therapy with AI now. I try to like reassure the AI... The generative imagery for slides was just not great until the latest image launch. And the text rendering of the images is so good that it completely changes the game.”
Slides and the image-gen question
Julius launched slides last month. The more interesting product decision sits underneath that: how to generate them. Sonwalkar says the right tool depends on what the user wants. For a single infographic or standalone graphic, GPT's image generation is now good enough, partly because text rendering has improved sharply enough to make it viable for presentation assets, not just creative imagery. For a full multi-slide deck where brand consistency matters across every page, Julius routes to a code harness, generating slides in JavaScript or HTML rather than relying on image generation. In practice, Julius picks the approach automatically depending on the request.
The distinction matters commercially. Image generation falls apart at scale across a deck, where inconsistency compounds. A code-generated harness holds style across slides in a way that image models currently can't. Sonwalkar frames this as a complement rather than a competition between the two methods, with image gen handling graphics and the code layer handling structure.
The roadmap, as described, points toward Julius becoming the place where data work ends rather than where it starts: analysis in, finished presentation out, tasks assigned, and results delivered without the user switching tools.
Every deal, every interview. 5 minutes.
TBPN Digest delivers summaries of the latest fundraises, interviews and tech news from TBPN, every weekday.