Bank of America sucks
Design your team for AI
If you love my writing :
I'm having a ball talking to people on the Life Sciences Today podcast. In the latest episode, I hosted Nate Beyor, Head of Life Science at Salt AI in Palos Verdes.
Last time I visited friends there, we ate at the Green Temple in Redondo Beach—a town stuck in a Beach Boys time warp of blondes, surfboards, and convertibles. The restaurant was founded by Elena Streskova and her husband, both passionate vegans. In 2000, it hit financial trouble and the husband ran off, but not before scrawling "Bank of America sucks" on the window. The food is still outstanding.
Salt AI is now working with the Ellison Medical Institute on AI for drug discovery—a gnarly challenge. Founded by Aber Whitcomb (founding CTO of MySpace), they've built a visual-first platform that gets workflows out of notebooks into shared environments. They run AlphaFold 2 at 22X speed and produce optimized Python for specialized hardware. No lock-in—bring your own code or eject theirs.
Their sweet spot? Data engineering leads who know how messy AI/ML workflows are under the hood. The ability to hide that complexity is massive. The marketing people have no idea.
Listen to the show here AI Workflows with Salt AI
Design your team for AI
How can we use AI purposefully despite the hype?
We are driven by knowledge—even as we chase hardware, humans, and pleasure.
Our personal growth is driven by learning.
Global economies run on design, engineering, software, pharma, logistics.
AI can accelerate all of it by an order of magnitude.
I don't say this lightly. I built an agentic AI system using LangChain, the X SDK, Sanity, and Python. Whether you're building AI for clinical decision support, protocol interpretation, or digital therapeutics—the challenges are deeper than code.
Read my essay The Great Agentic AI hoax to hear how I got my hands dirty.
You need to design your team for a post-AI age.
In this essay, I’ll discuss hype (is it bad?), why we code (how AI helps), why we are bad at code (so is AI) and how your team is your product (design for AI).
Hype (is it bad?)
Hype is strategic misdirection that creates urgency.
In mature industries like mobile, it's marketing misdirection. Like do you really need an iPhone 16 if you have an iPhone 14?
In innovative industries like AI, hype helps the whole industry reach escape velocity.
So hype isn't bad—but it comes with a catch.
You need to design how you use AI for your mission.
Kodak lost focus on their mission of photography. They focused on film, not photography. They missed the digital revolution and filed for bankruptcy in 2012.
Don't worship hype and AI tools.
Stay focused and design your mission to use AI properly.
Why we code (and how AI helps)
Joy
I've been writing code for over 50 years. There's a rush in discovering the power of a brain amplifier—having a machine run your instructions. For those who've written in dozens of languages, it's a complete joy discovering how much fun coding becomes with AI-augmented tools like PyCharm or Windsurf.
AI gives you an adrenaline rush of creating something fast with almost no effort.
Utility
As an undergrad physics student, I learned Fortran, Lisp, and SNOBOL in one semester programming languages course. Fortran and Lisp proved useful in my graduate solid-state physics research. Over my career, I've written hundreds of thousands of lines of code—there was always something useful at the end.
AI-augmented dev tools help you do useful things 6X faster.
But remember: code is about 15% of total effort. Most of the work in software development is in understanding the problem, designing solutions, testing them, and keeping them working.
Developing software is not shipping a product.
Shipping a customer-ready product is 3X the software development effort and requires packaging, documentation, support and validation.
This means that we're optimizing 5% of the problem.
Complexity
In one complex project, I led development of an airline reservation system for charter airlines. Locking seat inventory at scale is challenging—charter airlines can't use controlled overbooking like regular carriers. For scale, speed and concurrency, we wrote a lock function in assembly using XOR on shared global memory.
I'm now working on clinical data review systems. There's a lot of complexity in automating protocol understanding and executing workflows with human-in-the-loop.
I considered using a DSL to abstract away the complexity. I used OpenAI's o1 model to design the system architecture with a DSL. O1 was surprisingly creative designing the structure and syntax independently.
AI is surprisingly powerful for highly complex software design problems—actually more valuable than actual coding.
Learning
Every software project teaches you something: new algorithms, new solutions for old problems, old solutions for new problems. In designing the DSL, my mental model for data review process flow was a decision tree—humans handle trees better than graphs. Discussing this with o1, I learned about behavioral trees used in AI/gaming for character behavior.
AI is a valuable learning assistant.
Imagination
Programming activates imagination as you solve first-time problems, invariably interwoven with business requirements. In the charter airline system, we couldn't be sloppy with seat inventory—this pushed us toward the fastest possible strict concurrency control. For clinical data review, I imagined users implementing their own systems in human-like language.
AI is valuable as a dreaming assistant.
Why we are bad at code (and so is AI)
Precision
If one character is off, code won't work, may crash, or create security vulnerabilities. If requirements are wrong, code works but doesn't achieve business goals.
A clinical data management system we implemented was used successfully through multiple trial phases over 10 years. After FDA approval, the sponsor discovered that they didn’t have all the QoL data they needed for reimbursement codes—a major financial setback that one checkbox could have prevented by making the QoL form required.
Lack of control
You don't control the entire execution environment.
Here’s a real-world example.
One of our customers developed a wearable digital therapeutic controlled by a mobile app.
One day, they shipped a revision of their mobile software to the Apple App store with an error-handling bug on API calls that timed out (for example when the phone didn’t have connectivity).
When the error condition triggered, the mobile app went into an infinite loop, creating a DOS attack from dozens of patient endpoints to our cloud API services.
Dependencies
Python dependency hell is classic: Package A requires library X version 1.0, Package B needs version 2.0, and they can't coexist. Experienced Python developers have horror stories and preferred toolchains to minimize these issues.
Debug difficulty
Heisenbugs disappear when you try to observe them. Race conditions, memory corruption, buffer overflows work in debuggers but fail at runtime. Add timing-sensitive code, compiler optimizations, uninitialized variables, and hardware-specific cache coherency problems—debugging becomes hard and unfun.
Humans are bad at code because of bad requirements and poor understanding of other human needs.
So long as humans suffer from "what you did is not what I said but also not quite what I meant"
AI won't do better.
AI being good at code is an illusion holding only for well-specified, narrow tasks—not the messy reality of product development where requirements are fuzzy, changing, and contradictory.
The 800 trillion synapses in the human associational cortex do what LLMs don't: constantly integrating context, reading between lines, understanding unstated assumptions, cultural context, design goals, and navigating social/political dynamics that shape what software actually needs to do.
Your team is your product (design for AI)
A great programmer is a great coder, designer, and communicator.
Humans and AI are both bad at code.
Can AI augment humans in design and communication and help create great teams and products?
Great teams and products have conceptual clarity, good design, and simplicity, all a result of team collaboration.
Let’s examine these traits deeper.
Conceptual clarity
Conceptual clarity drives design, simplicity, collaboration, and team dynamics of your product.
Conceptual clarity is a “quality without a name”.
It’s connected to wholeness, grace, life, or rightness—but no single word quite captures it.
Let me illustrate with a product that lacks clarity - ConvertKit. ConvertKit is a mass mailer that does mass mail well. It has landing pages, calls to actions, sequences and more. The forms are beautiful.
But it's riddled with UX failures: poor database object defaults, hidden checkboxes, testing gotchas. The default for call-to-action pages is single execution, with the checkbox buried several clicks away. Users test repeatedly until satisfied, but it only works once—users will never know why the second test failed. Worse, their AI chatbot wasn't trained on current software versions, giving incorrect responses plus hallucinations.
There is a quality which cannot be named... a central quality which is the root criterion of life and spirit in a man, a town, a building, or a wilderness.
This quality is objective and precise, but it cannot be named.”
The search which we make for this quality, in our own lives, is the central search of any person... the search for that quality which has no name is the search for what is most alive.
– The Timeless Way of Building (1979) Christopher Alexander
(The book of design patterns that inspired object orient programming design patterns)
Design
I cannot force my team to do the things that I believe are right.
The only way I can encourage people to take action is to inspire them. The same applies to the team members.
Inspiration is not a side effect in design; it’s our primary responsibility and defines the way the software team works.
Design is often confused with nice looking things (like ConvertKit) but design is about not looking nice.
Design is a system for leadership. It combines analysis, visualization, decision making, prototyping and learning. Design uses all of our senses and especially our sense of time.
Design provides a path from a rough concept to a detailed solution.
Design provides a mechanism for separating signals from noise.
Design is leadership of the team that inspires others to lead.
Simplicity
In software, you know you're on the right path when lines of code start disappearing. When you arrive at the smallest program executing the solution, you've arrived.
Simplicity is fewer lines of code, smaller datasets, fewer forms. In clinical trials, collecting extra data often means missing the data you really need.
Collaboration
In my book "Ship Products People Love," I describe 23 anti-design patterns causing product team failures.
One is "Don't Flip the Bozo Bit"—once you write someone off as a fool, it's hard to undo. Curiosity dies. Collaboration collapses.
Team
Intimacy and trust seem to work best in groups of 5–7, balancing talents and personalities of developers, validation, product, UX/UI, DevOps.
Smaller teams have fewer interactions, less friction, and are faster and more productive.
With AI a small team can do an order of magnitude more.
Summary
Your human team is your product.
Especially in life sciences, where regulatory burden, distributed teams, and long timelines require intense trust and clarity.
The 800 trillion synapses in the human associational cortex do what AI can't: constantly integrating context, reading between lines, understanding unstated assumptions, cultural context, design goals, and navigating social/political dynamics shaping what software actually needs to do.
Design your mission to augment your team with AI. You can use AI to build more complex products faster.
🎙 Life Sciences Today — My Podcast for Leaders
I host leaders in life sciences. We talk about their personal journey.
I dive into the business model and how they capture value—not just the tech.
If you're changing the game in drug development, clinical ops, or RWD—I’d love to feature you.
👉 Think you’d be a great guest? Schedule a short call here I’d love to host you!
👉 If you enjoy reading this post, feel free to share it with friends! Or feel free to click the ❤️ button on this post so more people can discover my writing. 🙏
About Me
I’m a former pharma-tech founder who bootstrapped to exit.
Now I help TechBio and digital health CEOs grow revenue—by solving the tech, team, and GTM problems that stall progress.
If you want a warrior to work by your side, DM me on substack, LinkedIn or X