Discovery, leveled up
Why I'm building Coffee Run
Netflix changed how we find things.
Not perfectly, and not in every way — but the work streaming services did over the last fifteen years genuinely moved the field forward. We learned how to model content deeply (not just by genre, but by mood, pacing, the things people actually respond to). We learned how to model users beyond demographics. We learned how to surface the right thing at the right moment from a catalog of tens of thousands. Walk into Netflix today and the homepage knows something real about you. That wasn’t true twenty years ago. It’s a big deal.
The interesting question now is: where else does this paradigm apply?
Because those ideas — deep content understanding, deep user understanding, personalization as a way to bridge the two — aren’t streaming concepts. They’re discovery concepts. They work anywhere there’s more good stuff than any one person can sort through alone.
Yelp is the obvious counterexample. Yelp is fine. But Yelp tries to do everything: restaurants, plumbers, dentists, nail salons, auto shops, weddings. Dozens of jobs, none of them done especially well. Netflix works because it has one job: help you pick something to watch. The focus is the feature.
Coffee Run is my attempt to take the Netflix paradigm — deep content modeling, real personalization, discovery that feels editorial — and apply it to one small, specific job I happen to care a lot about: finding a great independent café.
It’s at coffeerunapp.com. Think Letterboxd, but for coffee shops. Log the cafés you visit, remember what you ordered, find new spots through people whose taste you trust. MVP is live and I’m building in the open from here on.
Why now
Netflix built its discovery engine the hard way — armies of taggers, years of A/B tests, custom ML infrastructure. That was the only path available at the time. It’s not the only path anymore.
The toolkit got dramatically better in the last two years. Modern LLMs can read a shop’s Instagram, their menu, a hundred reviews, and a few photos, and assemble a richer understanding of what that café actually is than a manual tagging team could produce in a week. That changes the economics of content understanding. A solo builder can now stand up the kind of catalog modeling that used to require a department.
I’m using AI two ways on Coffee Run, and both will be ongoing topics here:
Enriching the coffee scene itself. Every shop in the catalog gets pulled through a multi-stage enrichment pipeline — vision analysis on photos, structured extraction from first-party sources, a canonical layer of observations and intents that powers everything from search to guides to recommendations. The goal is to know each shop the way a thoughtful local would, at the scale of every city.
Shortcutting the product development lifecycle itself. Coffee Run is also a testbed for what an agentic PDLC looks like in practice — where AI is genuinely doing the work, not just suggesting it. From ticket specs through implementation through QA. I’m running a one-person team that ships at a pace I couldn’t have hit two years ago, and the lessons translate directly to how teams at larger companies (including mine) will operate by next year.
What I’ll write about
I lead product for personalization and discovery platforms at Crunchyroll — helping millions of anime fans find shows they’ll love in a catalog that’s growing faster than anyone can watch. AI is rewriting what that surface can be: search that understands intent, home feeds that adapt to who you actually are, recommendations that feel like a thoughtful friend made them. The day job and the side project are asking the same questions from different angles.
This newsletter will cover both:
AI-powered discovery and personalization — the paradigm, where it’s heading, where it hasn’t reached yet
The agentic PDLC — how product development changes when AI is in the loop, both inside large companies and on the projects we build in our free time
Coffee Run specifically — product calls, engineering decisions, shop visits, the things I get wrong
Some posts will be strategic. Some will be in the weeds on a specific build. Some will just be about a café.
Subscribe if any of that sounds like your kind of thing.
— Ben