Big shoutout to Rocket for partnering with us on Vibe Shift: The Builders Behind the Models — an evening where foundation models met product imagination.

This highlight focuses on a standout discussion between Harsh Verma and Johann Schleier-Smith, who unpacked what it really takes to move from powerful models to AI products that are scalable, usable, and built around real customer needs.

Their conversation brought sharp perspective to the core theme of the night: the builders behind the models are not just experimenting with AI — they are shaping how it gets deployed, trusted, and experienced in the real world.

Special thanks as well to Deepak Dhanak, Tanya Roosta, Puneet Gupta, and Francis Ma for adding thoughtful insights throughout the evening, and to Khyaati Brahmbhatt for helping bring the Twill and Rocket communities together.

Here are some photos from Vibe Shift: The Builders Behind the Models, capturing the conversations, connections, and moments that brought the event to life.

In case you missed it: the following article recaps the panel discussion with Harsh and Johann. You can still share your interest in a future session and get on the list for the next one here:
👉 Sign up here

One of the most grounded parts of the engineering panel came from Harsh and Johan, who shifted the conversation away from AI hype and toward a more practical question: what does it actually take to build systems that work reliably in production?

Harsh approached the discussion from a security lens, emphasizing that modern AI systems are no longer simple software workflows. They are increasingly agentic systems that can think, act, plan and operate across multiple layers, which means the old model of building first and securing later no longer works. His clearest point was simple: “now is the time to build things right.” For Harsh, that means thinking early about what data enters a system, where it flows, how models are being used, and what guardrails are in place before those systems scale.

Johann brought the infrastructure perspective. Speaking from his work at Temporal, he explained that as AI agents take on longer-running tasks across data processing, documents, and workflows, they need a more dependable foundation underneath them. Temporal’s role, as he described it, is to provide durable execution — infrastructure that helps make these systems “crash proof” so that if something fails, state is preserved and work can continue rather than restarting from scratch.

Together, their discussion pointed to a more mature way of thinking about AI products. Harsh focused on security, traceability, and deployment discipline. Johann focused on resilience, recovery, and the reality that AI systems need stronger operational foundations than a simple model wrapper can provide. One of Johan’s strongest insights was that AI engineering is really about getting “practically deterministic results out of non-deterministic technology.” That line captures the challenge perfectly: LLMs may be probabilistic, but production systems cannot afford to be.

The broader takeaway was clear. The next generation of AI products will not stand out just because they are capable. They will stand out because they are secure, observable, and resilient under real-world pressure. Harsh and Johann made the case that building serious AI products requires both: security by design and infrastructure that can recover when things break.

Subscribe to get access to future AI workshops and full replays of past sessions 👇

👉 https://luma.com/twill

Thanks for reading! Stay plugged into the latest founder news, events, and opportunities:

Let’s keep building the future, together. 🚀