Good Work People Fall Into an Obsessive Programming Hole

Maybe “obsessive” is too strong. Let’s call it “detail-oriented”?

“I have not failed 10,000 times—I’ve successfully found 10,000 ways that will not work.”—Thomas Edison

Hello! It’s been a little while since my last post. In fact, it’s been three months—long enough for my beloved (Cadbury creme eggs) to come and go, and for a batch of quarterly RSUs to vest. Sometimes “good work” means letting Substack marinate while I build the very things I want to write about.

So, what’s new?

I’ve been obsessively building an equity administration and accounting AI—one that’s actually useful in the real world. If you work in equity comp, you know there’s no room for “close enough,” not when people’s paychecks and company filings are on the line. My quest for getting this right means I’ve been learning (read: breaking and fixing) nearly every day since August.

Think of building an AI app as assembling IKEA furniture… but the instructions are scattered across five languages, three continents, and at least one dimension I’m not sure exists. I’ve finally embraced the joy of putting things together wrong, on purpose, just to see why it doesn’t work.

Let’s talk about me breaking the sidebar for a second.
If you’ve ever used a chat tool, the sidebar is where your chat history and saved answers live—the thing that helps you find your conversations (and sometimes your sanity). I’ve broken and fixed my version of that sidebar more times than I can count. I’m pretty sure “STOP BREAKING THE SIDEBAR” should be my new motivational poster. Or perhaps we need a “It’s been 0 days since AP broke the sidebar” sign.

In fact I think I’ve “failed” my way into becoming a mid-level full stack software developer:

• Built an AI-powered web app from scratch. I’m way past crafting the perfect equity expert prompts—I’ve got multiple AI models talking to each other, with a custom GPT picking the best tool for each job.

• Touched every layer: user interface (Next.js, React, TypeScript, Tailwind), backend (Supabase, Drizzle, Redis), authorization and authentication (Clerk—don’t ask me about NextAuth unless you want war stories), secure deployment, and the AI/automation side (Vercel AI SDK, OpenAI, Fireworks, Perplexity).

• Implemented Retrieval-Augmented Generation (RAG): Both the AI and users can teach the system new facts, which get stored as unique embeddings—think digital fingerprints for each nugget of knowledge.

• Architected a platform that balances privacy, security, and actual usefulness for people with real work to do.

What’s next?

The ideas list is long (and growing), but my most frequently requested? Make RSU releases less painful, without duct-taping logic on top of already-wobbly payroll and equity systems. Because not everyone wants to spend six figures a year just to get two stubborn systems to communicate.

If you’re still reading:
What do you wish you could automate? What’s your most soul-sucking process—the one you’d happily hand off if you were sure it could be done right?
→ Hit reply or comment here—seriously. I want to hear your horror stories (and your wish lists).

If you’re curious about early access to my new AI (her name is Izzie, more on that soon), reach out. Early users get to shape the tool and complain about me breaking things in real-time.

Here’s to a new season of learning, building, failing, and building better!

“Failure isn’t failure. It’s experience.”—Daymond John (Shark Tank)