AI Across The Product Lifecycle Podcast

The episode contains solid practitioner insights about AI's impact on engineering workflows, particularly around development velocity and organizational changes. However, significant portions consist of soft setup questions, repeated themes (cost as token-based like AWS, AI everywhere), and circular discussions without novel depth. The most concrete insights - like moving testing/QA to bottleneck, using AI agents through Slack, orchestrating cheap vs. frontier models - are valuable but not densely packed across 56 minutes.

“we're hiring a lot of kind of people in the back end, right? So like have Claude code or cursor, right? Like write up all of these new features. It's to a point where we have cursor hookup to Slack. And so when a customer reports a new feature, I can basically forward it to cursor and say, cursor, go do this.”

“our focus has really shifted. We actually find like kind of testing and like QA and like validation to now be our bottleneck.”

The episode largely recycles established narratives: AI democratizing engineering (parallel to text-to-code), trust-but-verify frameworks, cost being analogous to AWS token-based pricing, and the organizational readiness gap. While the guests add perspective as practitioners building in the space, the core ideas are not contrarian or first-principles. The discussion of orchestration between model tiers is somewhat fresher, but framed as an obvious operational necessity rather than a novel insight.

“it's kind of like a crawl walk run approach and it's like trust but verify is like our thesis, right?”

“we're going from, you know, headcount costs to to maybe tokens, right? You know, when we did AWS, you're going from physical real estate costs and hardware costs to cloud computing. And and that you could also say is kind of like a token”

Both guests are relevant practitioners: Pradyut is co-founder of Build, a CAD data management platform with real customer traction; Martin Bielicki is CEO of Bench, an AI orchestration layer for engineering. Both have shipped products and have direct experience with customer adoption challenges and AI integration at scale. However, neither is from a dominant market player (e.g., Tesla, SpaceX, Ansys, Autodesk), and the companies are early-stage, limiting their ability to speak from massive-scale transformation experience. They are serious operators, not thought-leaders, but operating at moderate rather than exceptional scale.

“Build is a CAD data management platform. We're essentially connecting all things engineering into manufacturing. Our focus is really around data management and automating the workloads built on top of that that data.”

“at Bench we're building an orchestration layer for engineering. So essentially we're looking to apply AI to automating AI end-to-end workflows in engineering, so spanning CAD, simulations, PLM and the like.”

The episode includes some concrete examples: cursor/Claude hooked to Slack, concept-to-delivery in one day, hiring bottleneck shift to QA/validation, use of Graphite for testing. However, much of the discussion remains abstract: vague references to customers, no revenue figures, customer names redacted, no specific timelines for product milestones, and few quantified impact metrics. The CAD/simulation/PLM workflow discussion lacks concrete example use cases with measurable outcomes.

“we have cursor hookup to Slack. And so when a customer reports a new feature, I can basically forward it to cursor and say, cursor, go do this. It'll build it in the back end and it's all interacting through Slack. And then we essentially have Graphite, which you guys probably know about, which will do some of the testing.”

“from concept to delivery in a day, which was obviously never possible before.”

The host asks some probing questions (cost curve, prompting vs. autonomy, merge conflicts with multiple agents) and demonstrates genuine curiosity, particularly around technical debt and organizational friction. However, follow-ups are often soft and don't press when guests deflect (e.g., Pradyut says his CTO should answer on merge conflicts, and Fino lets it drop; Martin similarly defers). The host also gets sidetracked into personal anecdotes (GSD, personal AI spending) rather than deepening guest answers. Some questions are conversational setup rather than substantive probes.

“To what degree are you depending on prompting skills as opposed to letting AI and fill in all the gaps? To reframe it slightly, how do you see the friction curve trending for AI in terms of autonomy?”

“So if you have multiple agents all hitting your code base at the same time, they could be r overriding each other and agile was supposed to help us not be destroying somebody else's code. So how do you how do you manage that?”