I built a daily GitHub AI leaderboard — 3 things nobody tells you about star counts

I'm a solo developer in China. In August I got annoyed enough by AI tool directories to build my own, and a couple of months later the weirdest thing happened: the biggest chunk of my readers is in India, Pakistan, and Bangladesh. Not the US. Not China.

This is a build-in-public post about how the site works under the hood — the leaderboard pipeline, the star-count math, and what the data actually taught me. I'll be specific because vague "I built a directory" posts help nobody.

What I ended up building

ai-all.info is a bilingual (English/Chinese) AI resource site. The parts that took real work:

• 1,666 tools across 35 categories — hand-curated, not bulk-imported from some other directory's API
• 590 English model reviews — one page per model/tool: what it does, pricing, how it installs
• A daily GitHub AI leaderboard — the part I'm actually proud of, detailed below
• llms.txt — a machine-readable site map for AI agents, with robots.txt tuned per bot (GPTBot and ClaudeBot allowed, Bytespider blocked)

The boring stack: Next.js + Supabase, GitHub Actions building a Docker image pushed to GHCR, docker compose up -d over SSH to a cheap VPS. The whole infra costs less than a conference lunch.

The leaderboard: stars gained today, not total stars

Every "trending AI repos" list you've seen ranks by total stars, which means a 2023-era project sits at #1 forever and a genuinely hot new agent framework drowns at #80.

So the daily board ranks by stars gained that day, with separate views for overall hot repos, tools, skills, and raw star counts. The pipeline is deliberately small:

1. A nightly Python job scrapes GitHub's trending pages (Python, and a few other languages), then parses each repo's "N stars today" badge directly off the page
2. An AI keyword filter (a hand-tuned include/exclude list — llm, agent, rag… in; iptv, home-assistant… out) decides what's actually AI
3. The result is written to latest.json; before overwriting, the previous day moves to archive/YYYY-MM-DD.json — currently 81 days of history
4. An automated commit (chore: auto-update AI GitHub daily leaderboard) lands on the repo every night; the site renders from that JSON

Thing nobody tells you #1: GitHub already computes "stars today" for you — and you inherit its opinions. That little "N stars today" badge is GitHub's own number over GitHub's own time window. Your leaderboard is only as good as trending's own selection. The badge says nothing about repos GitHub's algorithm didn't pick to show, and the window isn't midnight-to-midnight in your timezone. I stopped pretending my board is the objective truth about star growth; it's my snapshot of what GitHub surfaced that day, which is still useful — but you should know what you're actually measuring.

Thing nobody tells you #2: you need the archive from day one. GitHub gives you today's badge and nothing else. The only way to answer "is this repo still growing?" or "how hot was last Tuesday?" is your own daily snapshots. My archive JSONs are what makes the tools/skills/stars views meaningful — and yes, a partial scrape (a page fetch fails, a keyword filter misjudges) is only visible because I can diff against yesterday. Without the archive, every trend claim on the site would be hand-waving.

Thing nobody tells you #3: the ranking changes what people notice. When the board ranks by daily-delta instead of total-stars, the #1 slot rotates between agent frameworks and infra tools within days of them blowing up. Example from this morning: NVIDIA/auto-ontology — a semantic-graph reasoning framework — tops the star-surge board with 1,155 stars gained in a day, while stablyai/orca (a fleet-of-coding-agents workbench) sits #1 on the hot board. On a total-stars list neither would be anywhere near the top this week. That rotation is the whole point: the site tells you what AI builders paid attention to this week, which no static directory does.

What the analytics surprised me with

I launched expecting — honestly hoping — for US traffic. What Search Console actually shows:

• India is the single biggest click source — about a third of all clicks — and South Asia as a block (India, Pakistan, Bangladesh, Nepal) is close to half. English queries, from South Asia, on a site built by someone in China. The "developer Googles the exact model name at 2am and lands on your review page" pattern turned out to be global, not American.
• Long-tail model-name queries are the actual entrance. Not "best AI tools 2026" (a keyword owned by sites with 100x my domain rating). People search belin doc, dramaclaw tutorial, holopix ai — product names and one-off queries no keyword tool surfaces. One of my proudest pages is the first DramaClaw tutorial on the internet — written the day the app hit GitHub trending, because I knew the official README wouldn't explain installation pain points to a non-technical user. Nobody else had it. Google had no choice.
• Freshness beats authority, at first. The daily-updated leaderboard pages earn internal links from being the most useful thing on the site, and they get crawled faster than my static pages ever did.

What I'd tell another solo dev building in this space

1. Don't build "another AI tools list." 1,600 links you didn't verify is a worse version of what exists. The defensible position is data nobody else keeps — in my case, the daily archive.
2. Write the llms.txt before you need it. AI assistants and crawlers are a real referral source now, and a well-formed llms.txt costs you 30 minutes. It also makes your content the thing that gets quoted when someone asks an LLM for recommendations — the best distribution channel a small site can get.
3. Automate the boring pipeline completely. The nightly commit has run unattended since August because it's the same one job in one repo, and it snapshots before it overwrites. A leaderboard that lies once dies.
4. The market you expected is not the market you get. My Chinese pages get a trickle; on the English ones, India alone out-clicks the US by nearly 3x. If your content is genuinely useful, English is a bigger lever than anything else I tried.

Numbers after 11 weeks

1,666 tools, 590 English reviews, 81 days of leaderboard archives, a handful of tutorials, one VPS, and zero ad spend. The board updates every night whether I'm awake or not — which, for a solo project, is the closest thing I've found to job security.

If you want to see the pipeline's output for yourself, the daily leaderboard is public and the archive JSONs are what the charts are actually made of. And if you build anything similar — I'm genuinely curious what your "thing nobody tells you" list looks like.