Five Machine Learning Projects I Wanted to Work On — and What Actually Happened

Back in May 2017 I published a short post on Medium listing five machine learning problems I was itching to tackle. I was deep in the streaming video world at the time — running a multi-channel network on YouTube, managing a massive Latin American content library, obsessing over subscriber funnels. More product manager than coder, and ML felt like a superpower I could almost reach but couldn’t quite grab.

Nine years is a long time in this space. Here’s what happened.

1. Analyzing a humongous video library

“I have over 17,000 hours of broadcast quality video… I’d love to make compilations of the best scenes, but looking for them is a daunting task.”

Built it. Killed it. Learned from it.

I built this. Scene detection, metadata tagging, making 17,000+ hours of broadcast content searchable. The pipeline worked. The problem was that processing that much video with the tools available then was slow and expensive — too slow and too expensive given what we could actually spend. We shut it down.

The lesson wasn’t that the idea was wrong. “This works technically” and “this works for our business right now” are two separate questions, and I’d only answered the first one. We were early on the cost curve, not wrong on the direction. That distinction matters when you’re advising clients who want to build something similar today — because the cost curve has moved a lot.

2. A recommendation engine for video viewers

“Imagine if you took each user’s watch history, search queries, added Wikipedia and IMDB, maybe looked at the campaigns that brought each user in…”

Directionally right. Harder than it looks.

Personalization at scale is table stakes for any streaming platform now. Off-the-shelf recommendation infrastructure exists. But I’d be doing a disservice to the engineers who actually built these systems if I said the problem is solved. Cold-start — what do you show a brand new user with no history? — is still hard. Feedback loop bias is still hard. The tooling got much better. The nuance didn’t go away. What changed is that teams can spend their time on those problems instead of rebuilding plumbing from scratch.

3. Analyzing onboarding funnels

“When you’re paying top dollar to acquire customers, you need to make sure they’re not leaking out of your funnel… you want your analysis to scale.”

The tools arrived. The behavior didn’t.

Churn prediction and lifecycle scoring are in reach for almost any company now. But the gap I keep seeing has nothing to do with the model. The data lives in one team, the decision to act lives in another, and nobody owns the moment between insight and doing something about it. Most companies already have signals telling them a user is about to leave. They just don’t act on them.

4. Pattern-finding in a multi-channel network

“I’d like to sit down, look at the data we have and look for patterns that indicate potential success — not the next viral hit, but sustainable growth.”

Needle found. Wrong haystack.

Finding signal across thousands of channels is something AI handles well now. But the real lesson from that period had nothing to do with ML. When your business runs on someone else’s platform, their algorithm rewrites your economics overnight. YouTube changed how it distributed content, how it paid creators, what formats it rewarded — and the MCN model buckled. No data analysis was going to save that. You have to ask the strategic question before the technical one: what ground are you building on, and who controls it?

5. The open question

“What else do you think I should work on? What other cool ML opportunities do you see?”

Still the only question that matters.

I ended the 2017 post admitting I was more PM than coder, and that building any of this might take longer than I’d like. Tools like Claude Code, Codex, and Gemini have changed that equation. A working prototype is now hours away, not months, even without a dedicated engineering team. The implementation gap I was describing has mostly closed.

Which means the scarce thing shifted. It’s not building anymore. It’s knowing what to build. Finding the right problem, understanding what’s actually constraining it, translating a messy operational reality into something a model can usefully touch — that’s where the work is now.

So I’m back at the same question I asked nine years ago. The 2017 list is mostly retired. Time to build a new one.

Looking back, what strikes me is how stable the underlying problems were. Retention, discovery, scale — same issues, just more tools pointed at them. The speed of AI development has made the question of which problem more important, not less. Anyone can spin up a model now. Figuring out where to point it is still the job.

What’s on your list? If you’re sitting on a problem that feels like it should be solvable by now — a workflow that’s still manual, data you’re not acting on, a library of content or customers that aren’t working as hard as they could — let’s talk.

Originally published May 2017 on Medium as “Five Machine Learning projects I want to be working on now.” This is the 2026 update.