Show Notes
This sounds like a silly click bait joke, and it kind of is. It’s also just how Aravindh works - ever since I’ve known him (close to 20 years at this point) he’s had this intense work habit. I admire how he does it; he knows not to get too far ahead of his team mates, for instance - so me might spend 20 hours learning something new for fun, 20 hours on a side project (like the algorithmic trading bot he built), and only work 50 or so hours a week or so on his main job. All of this to say, setting down tools that make you arguably more productive 28% of the time is showing intense restraint - and I admire this about him, too. Aravindh swears by it - it helps him have a much better sense of the codebase that he and his GenAI tools are making. It also helps that he’s one of the best software developers I’ve ever worked with. He balances confidence, expertise and the openness of the “beginners mind” like no one I’ve known. The other thing he talked about in our interview makes new use of an existing axiom - that you should “own the seams”. In this interview, Aravindh argues that we should prioritize working closely with our GenAI tools to define REST interfaces, service contracts, classes, interfaces, traits and so on, and exert slightly less control over how it works between these seams. “Own the seams” often gets used when talking about systems integration; that if you use IBM’s Websphere Commerce (which has the shopping cart) and also Sterling’s Order Management System (which manages the order that the cart becomes), try to not let these two important packaged systems to talk directly to another. Even placing a thin shim of a proxy between them begins to give you some control - and better yet, place some kind of adapter between them to prevent Sterling from becoming aware of or dependent on Websphere Commerce’s internal domain jargon and vice versa. This particular use of owning the seams flips it on its head a little bit - instead of enterprise software teams placing domain facades between two massive packaged software systems in order to have more control, its you exerting control over your GenAI coding tools in order to have more control over your code. Spoken like someone who spends 28% of his week working in his code base by hand! As always, thanks to our Founding Sponsor, Navalia! The “long working hours” part of this story definitely deserves a “don’t try this at home” warning - it won’t work for everyone. I do think it works for Aravindh, though.
More from Engineering with AI
Ian Johnson is Staff Software Engineer at Parento and author of Harness Engineering. And, he's the first guest that talked to me about how they are building a factory. Ian's book, Harness Engineering: https://leanpub.com/harness-engineering
In this episode, Pete Hodgson discusses the evolving landscape of AI-assisted engineering, the importance of governance in AI workflows, and how organizations can empower grassroots innovation while maintaining control. He shares insights from his role at Tribe and explores the future of agentic workflows and token economics.
Hari Krishnan is back to talk about advancements in spec driven development, about advancements in open weights models, and about the book he's writing for Manning. You can get the book for 50% off (and possibly save even more this week?) with the Manning Early Access program here: https://www.manning.com/books/spec-driven-development You get access to the first three chapters, and the full book once its ready.
Asoba’s CEO Shingai Samudzi is no stranger to ML, as a data scientist he’s put it to use throughout his career. Now, as Asoba looks to accelerate solar energy in South Africa, he’s assembling a new approach; combining a per inverter JEPA world model with Nahenda, a model they’ve developed based on Qwen fine-tuned specifically for epistemic behaviour. Though this would be cool enough as an experiment, the early results are promising. On recent runs analyzing actual solar inverter data, they more accurately spotted problems (and ignored false positives) with equipment. In solar power, this improves margin, and can make or break a solar plant. https://huggingface.co/asoba More of Shingai’s research More on SEP, Specification enhancement proposals
Rob is the CTO at Circle CI, and I was really interested to hear what all he can see from their vantage point - and it didn’t disappoint. He had lots of broad ranging insights in to what teams are doing all over the capability map; median to P90, how they’re different and what’s working. A super insightful conversation! And he’s also just got a great energy, super fun conversation. Key Topics AI’s impact on software development and testing Organizational adaptation to rapid change with AI Tools and mental models for AI-driven coding Revising requirements and workflows in the AI era Merge efficiency and reducing friction in CI/CD pipelines The role of agents and prompts in software engineering Organizational hiring and contribution patterns in AI-driven environments Impact of AI on software testing and debugging Shift towards product thinking and distributed decision-making The role of tooling and automation in quality assurance Changing team structures and roles in software organizations Strategies for managing system complexity and incident response As usual, thanks to Navalia our founding sponsor! We have an embarrassment of riches when it comes to deciding what will be next, I have interviews ready from so many super interesting guests that I can’t decide who will be next yet: Maybe Pete Hodgson; our second ever guest, super insightful and in a new role where he’s seeing even more, back to talk about what he’s been up to Or Shingai Samudzi, CEO and Founder of Asoba about how they’re using JEPA world models to improve solar in Africa or Ian Johnson, author of https://leanpub.com/harness-engineering who has been building agent factories at Parento? Plus, I have even more interviews scheduled! It’s been a great summer and its looking like it will be a fantastic fall on Engineering With AI.