Personal Software: Kevin Rose Built His Own Techmeme for $300

Kevin Rose built Digg, an app that hit 38 million monthly users. He co-founded Zero, a fasting timer doing double digit millions in revenue (his words). These days he's a partner at True Ventures and rebooting Digg with Alexis Ohanian.

So what does a guy like that build at night, just for himself?

A news aggregator he may never launch. Total build cost: about $300 in AI credits.

This is the best case for personal software I've seen all year. Kevin went on Greg Isenberg's podcast, shared his screen, and walked through the whole thing: the stack, the pipeline, the scoring engine, the philosophy. There's no launch plan and no pitch deck, just a founder building the exact tool he wants to exist.

Watch it, then I'll break down what he built and why it matters for you.

What Kevin Actually Built

The project is called Nylon. It's his nights-and-weekends incubator name, the thing he codes on when the house goes quiet.

Nylon is a Techmeme-style aggregator for AI news. Kevin is upfront that it's not a competitor to Techmeme (he has huge respect for Gabe Rivera, who's been building it since the Digg days). He built it to answer one question: can I, alone, build something on par with Techmeme in about a week?

The answer, watching the screenshare, is yes. And that answer should mess with your head a little.

Here's what's under the hood:

• 63 sources feeding in: RSS feeds, Reddit, TechCrunch, Mac Rumors, The Verge, plus dynamic scrapers for sites that broke their RSS
• Every article lands in a Postgres database and gets enriched: title, description, author, full body text
• Articles get clustered by topic, so 47 stories about the same Nvidia investment collapse into one card
• A scoring engine ranks every cluster by impact, novelty, and viral potential

In the 24 hours before the recording, the system had enriched 2,284 of 2,288 incoming stories. That's 99.8% of a real news firehose, processed by one person's side project.

Context makes this wilder. Kevin's day job right now is the Digg reboot with Alexis Ohanian, where Justin Mezzell runs the day-to-day as CEO and Kevin pushes on experiments in the labs area. Nylon is where he tests the edges.

The $300 Stack, Piece by Piece

This is the part worth stealing, because the whole pipeline is cheap, boring plumbing you could copy this weekend.

When an article comes in, three services race to fetch the best version of it:

• Iframely pulls rich metadata: title, description, best image, embed codes
• Firecrawl crawls the full article body, with stealth modes for picky sites
• Gemini is the last resort: when both fail (Reddit blocks everything), he tells Gemini to turn on search grounding and figure out what the story is about

Then an AI judge compares the three results and picks a winner. On the example he showed, Firecrawl got a 403, Iframely came back low quality, and Gemini won the main content. 353 stories needed that Gemini fallback in one day.

For summaries he uses GPT-5 mini ("fast as hell and cheap"). For search he hits the Tavily API: when three or more sources cover the same story, Nylon goes hunting for coverage outside its own feeds. Embeddings run on OpenAI's large model and get stored right in Postgres with the vector extension.

The glue holding it together is trigger.dev, which runs his background jobs as durable cloud tasks with free retries. Kevin runs thousands of tasks a day through it for under $100 a month. Greg points out it's open source too. (Vercel's new Workflows beta does something similar for free if you live in that ecosystem, and Kevin recommends Vercel's AI Gateway so you can swap models without recoding.)

Add it up: a failover content pipeline, vector search, topic clustering, and an orchestration layer. Roughly $300 in AI credits to build, double digits a month to run while he tinkers.

Five years ago this was a seed round and a team of four. I wrote about this shift in my vibe coding workflow breakdown, and Kevin's build is the most advanced version of it I've seen.

The Gravity Engine Is the Actual Genius Part

Aggregating news is a solved problem. Ranking it by what matters to YOU is not.

So Kevin built what he calls a gravity engine: a panel of AI judges that votes on every story cluster like an editorial board. Each story gets scored on dimensions like industry impact, consumer impact, actionability, technical depth, and entertainment value.

On the Nvidia and CoreWeave story he showed, the board voted: 88% impact, 86% intellectual gravity, 75% early trend, 60% viral potential. Total editorial score: 85%.

The dimension he cares about most is novelty. His example stuck with me: the first time Bitcoin showed up on Hacker News, everyone called it stupid. A novelty detector tuned to a long horizon would have flagged it off the charts. As a builder and an investor, that early weird signal is exactly what he wants surfaced.

And then there's the unsettling party trick. Because every article becomes a vector, Nylon can measure how similar supposedly independent articles are and when they published. When five "reviews" land within an hour of each other hitting identical key points, that's an undisclosed paid campaign. Kevin says he's started catching them. ("It's pretty effed up.")

One guy on nights and weekends built a sponsored-content detector the FTC would envy.

Why He Might Never Launch It

Here's the twist that makes this a personal software story instead of a startup story.

Kevin might never ship Nylon. And he's completely at peace with that.

His argument: we're entering an era where, if a workout app doesn't track the one metric you care about, you just build your own. The question stops being "is there a market?" and becomes "how many people are like me?" Maybe a thousand. Maybe 100,000. Maybe just you.

Two lines from the episode I keep thinking about:

→ "The future developer, it's not going to be what you build as much as what you don't build." When building anything costs almost nothing, taste and subtraction become the skill. He said the perfect version of Nylon would cut 90% of its features.

→ If you have 500 people who really love what you made, that's 500 people outside your house cheering. We've lost perspective on what success means.

He also torched the standard objection to vibe-coded apps. Yes, they can be buggy and fall over under load. Kevin's response: those are great problems. The hard part is building something people actually want. If 50,000 people are beating your door down and the server crashes, you can hire engineers for that.

And if nobody wants it? His math is just as clean: "It cost me $500 and I flush it down the drain and onto the next thing."

That's the real unlock of personal software. The cost of being wrong dropped so low that every idea deserves a cheap test.

The Fun Projects Keep Becoming the Big Ones

There's a pattern in Kevin's career that he admits he can't explain.

Digg started in 2004 as "I just want to see what happens when people vote on things." Eighteen months later, 38 million people a month.

Zero started as a silly timer to track his fasts. It turned into a company doing double digit millions in revenue.

Both were toys first. Businesses second.

He's still running the same play. On the episode he demos a second night project: a blog where a blurred live video of the author plays in the background, so you can feel a human presence on the page. He posted the idea 12 years ago, when real-time video compression in the browser didn't exist. Now he rebuilt it with Claude Code in a weekend and plans to open source it. No business model. Just fun.

Greg named the pattern on air: the projects you put out just for fun somehow end up becoming the biggest businesses. Kevin's answer: "You are 1,000% right. It is the weirdest thing that I do not know how to explain."

One more detail that got me. Kevin was a computer science dropout who could grasp every concept but couldn't retain syntax, and only six months ago learned why: he has aphantasia, no mind's eye at all. People told him to picture sheep jumping a fence and he thought they were joking. AI now fills in that gap completely. Greg dropped out of CS too, three classes short, because he could write 92% of the code and never get it to compile.

Two CS dropouts, now shipping production systems solo. The excuses are gone.

What This Means for You

You're not Kevin Rose. You don't have his network or his True Ventures checkbook. Fine. The playbook still transfers, because none of it needed either.

→ Build your own tools first. The thing that annoys you daily is a weekend project now. Mine run my whole content pipeline, and like Kevin's local AI setups I covered here, they cost almost nothing to operate.

→ Use boring plumbing. Postgres, a queue with retries, three content fetchers and a judge. Kevin's architecture is copyable in an afternoon of prompting.

→ Score your inputs. The gravity engine idea works for anything: leads, feature requests, content topics. Have AI judges vote on dimensions you define.

→ Let toys be toys. Ship the fun thing without a business model attached. The track record of "just for fun" is suspiciously good.

One more thing from the episode worth repeating. Kevin, an actual VC, said most founders shouldn't take venture money. A business doing $2M to $10M a year that you own 100% of is a win, not a consolation prize. When a VC tells you that on camera, believe him.

(He also offered free desks at his new Venice studio in LA to anyone building cool AI stuff. DM him on X. Not a bad trade for a plane ticket.)

FAQ

What is personal software?

Personal software is software you build for yourself first, shaped to your exact workflow, with no launch or revenue requirement. AI coding tools dropped the cost of building it to near zero, so founders like Kevin Rose now build production-grade tools (like his Nylon news aggregator) as nights-and-weekends projects for a few hundred dollars.

What is Kevin Rose's Nylon project?

Nylon is Kevin Rose's private AI news aggregator, shown publicly for the first time on Greg Isenberg's podcast. It ingests 63 sources, enriches articles through Iframely, Firecrawl, and Gemini, clusters stories with vector embeddings in Postgres, and ranks them with an AI editorial board. He built it alone for about $300 in AI credits and says he may never launch it.

What tools did Kevin Rose use to build it?

The stack: Postgres with the vector extension, Iframely and Firecrawl for content fetching, Gemini with search grounding as a fallback, GPT-5 mini for summaries, OpenAI embeddings, the Tavily search API for expanding story clusters, and trigger.dev for durable background jobs (thousands of tasks a day, under $100 a month). He codes with Claude Code and recommends Vercel's AI Gateway for swapping models.

Is vibe-coded software good enough to launch?

Kevin's take: the bugs-and-scale objection is backwards. Finding something people want is the hard part. If your vibe-coded app crashes because 50,000 people showed up, you can hire engineers to fix that. Start scrappy, prove demand, formalize later.