I built a site that estimates what your repos could be worth to AI companies

Pietro — Thu, 22 Oct 2026 08:04:37 +0000

I keep an eye on companies that train AI models. I spent the past year or so creating the best AI training / data annotation job aggregator out there, and the natural next step was to expand to enterprise data licensing.

Data labs are also interested in licensing private code, and some publish what they pay per token: one rate for source code and a lower rate for Git history. That gave me enough to build a calculator. Then the calculator turned into two sites.

Githubworth.com estimates what the repos on a GitHub profile might sell for based on real rates from active buyers. It's a spin-off utility tool from Dataconto, which helps companies license their operational data and coding bundles to data labs. It shows which AI companies buy code or company records, such as email, chat, tickets and docs, and what they ask for.

Both sites use the same pricing package and run on Cloudflare’s free tier

How the price is calculated

A repo’s estimated value is:

source tokens × $100 / 1M + history tokens × $40 / 1M

Those are published per-token rates for non-exclusive licenses. The methodology page links to the sources and notes when each rate was last checked.

The counter follows the published rules buyers use:

Code files count in full.
Docs, config and data count up to 10 times the size of the code.
Vendored folders, bundles and minified files are skipped.
Tokens are estimated as bytes divided by 4.
The counter runs in your browser, so folders you drop into it stay on your machine.

Estimating Git history
GitHub shows a repo’s languages and commit count, but not how many tokens are in its history. To estimate that, I cloned 532 public repos across 18 languages, counted their full histories using the same counter, and fitted this formula:

`history ≈ 45 × source^0.48 × commits^0.88

Findings

History grows almost one for one with commits, and roughly with the square root of source size. That fits how Git stores changes: each commit records new copies of the files it touched.
Once source size and commits were included, repo age, contributor count and repo size didn’t improve the estimate.
For the median repo, history made up 80% of the estimated value.
The estimate can be off by about 1.8x, so the site shows a range. The low end is 0.27 times the estimated history, and the high end is 3.8 times. That range covered the counted value for 77% of the repos.

Repos with 10 commits or fewer are the toughest to estimate. For those, the site shows only the range.

I also tested five trending repos that weren’t in the sample. ComfyUI came out at $7,448 estimated, compared with $7,112 counted. All five results fell within their ranges.

Limitations
The estimate relies on a small number of published rates. Most companies review each codebase and set a price individually, without publishing a rate. They aren’t included in the formula.

There are two other important limits:

Buyers review code before paying, so this is a list-price estimate, not an offer.

Nobody licenses public repos. For a public profile, the number shows what that much code and history might list for if it were private.

Designing the best AI Training Jobs Aggregator and standing out from copycats

Pietro — Wed, 21 Oct 2026 08:24:19 +0000

I'm a UX-researcher-turned-indie-hacker, and for years I've built side projects on my own. Somewhere along the way I stopped doing what I actually enjoy (the research bit) while entirely focused on coding. In an attempt to get back to my roots, around December last year I started applying for UX research roles at Mercor, one of the platforms that pay experts to help train AI models. I did a bunch of interviews and then waited endlessly for opportunities to become available to me.

While I waited, I spent a lot of time comparing different AI training and data annotation platforms, and realised there wasn't a good enough aggregator out there.

In fact, most if not all the job boards and aggregators in the space were obviously AI generated "slop" that looked like the same template with a different logo and terrible copy that made no sense and had a penchant for em dashes.

This is how I ended up creating aitrainer.work, the UX-first platform for AI trainers and job seekers in this space.

From basic job board to all-in-one resource hub

What started out as a basic job board has evolved into a multi-faced platform for job seekers that includes a Talent Pool (from which I manually match candidates with the newest and best suiting job opportunities based on their skills and background) and a free Academy with modules tailored around real tasks from the data labs and module content constantly improved based on user feedback. When a user requested flashcards for one of the modules, we added them within a day, then brought them to all other modules as well.

We also have guides, news, mockup interview questions, practice tests and a certified qualification that people who pass the exam can put on their CV and linkedin.

Tech stack

The site runs on Astro and Cloudflare Workers, with Cloudflare D1 as the database. Custom scrapers read each platform's job board every day.

I built the first versions with Amp, an AI coding agent by Sourcegraph, and switched to Claude Code in April. Vibe coding is honestly such a powerful tool for a UX person that's self-admittedly not that good at coding but knows exactly what makes a good product. I can focus my creative juices on UX and service design while letting my agents deal with the building.

That's not different from when I was working for a product team and my role was to tell other people that their ideas were bad, prove it with science (usability testing) and giving them recommendations on how to improve our product. Finally, I am back to my roots!

aitrainer.work vs Other AI Training Jobs Aggregators

I genuinely believe that aitrainer.work is the most user-friendly AI training job board out there, in fact, I believe it's the only platform that is actually designed around the users: every design decision is made based on data collected and insights generated from user research, and product has been constantly improving and iterating based on user feedback.

I also aggressively focus on making my site look as little AI-generated as possible. Yes, I may be using Claude for coding, but if a user lands on any page and immediatedly thinks "an AI made this", that's a failure in my mind. I am especially obsessed with copy: working with UX copywriters in the past, I know how crucial it can be to any product, and I also know that AI agents are notoriously bad at it.

I pride myself on making aitrainer.work SOUND much more "human" than other job aggregators who rely exclusively on AI for every aspect of the design, including and especially user-facing copy.

I decided to curate less platforms and job listing spefically to ensure not to create poor quality listings en masse, and also to filter out "spammy" or dubious platforms and only focus on quality players. For example, we do not list platforms that pay their trainers with their own cryptocurrency, or that have overwhelmingly negative feedback from workers. Other aggregators may list a larger number of job opportunities, but we go for quality over quantity.

Hiccups and future plans
The site has been live for almost a full year now and I would deem it a success, with generally positive feedback from users, over 1700+ people in our talent pool and 400 successful hires.

Yet, it has been facing some SEO/indexing issues and a recent surge in copycats and competitors I have had to dealt with. I am battling this by enhancing the site's own proprietary data, such as our successful placements track record and open data with unique stats and research insights nobody else out there can replicate.

I have also been experimenting with gamification and giving people a reason to come back to the site on a daily basis, while practicing their RLHF skills.

Finally, I am also working on not one, not two... not three... but four spin-off platforms, all with different target users, and hence each with their own design and revenue models, which has been keeping me busy but hopefully will pay out and ultimately strengthen and diversify the whole product portfolio.