Author here. I've shipped multiple Next.js projects to production with LLMs writing over 90% of the code, and this post outlines a project structure that aids LLMs and doesn't trip them up.
TL;DR: Keep it flat, obvious and simple. 7 top-level folders (app, components, database, emails, functions, lib, public) instead of deep nesting. Explicit server/client separation. Everything has one obvious place to live.
Happy to answer questions about working with LLMs on larger codebases or specific patterns I've found useful.
I created this template because I couldn't find any resources to run ClojureScript on Cloudflare Pages (one of the fastest and cheapest deployment options available today).
We do have a checkout flow, and yes, that page is visited by a very small number of users. However, you don't actually think I deleted the checkout flow, do you?
Our situation is a bit unique, because we're currently a two person company. I'm familiar with the product, our users, and the codebase, which allowed me to move quickly. This probably wouldn't work at a larger organization.
I didn't say I deleted everything that had less than 0.1% of page visits. :)
For instance, I didn't delete the "Account Settings" page. Features don't need a high percentage of traffic to be important, but less important features generally account for a very low percentage of traffic.
I've worked on the codebase since day one. We structured it carefully so that modules didn't depend on one another. All shared components were in a separate folder. The structure of the web application lends itself naturally to such code organization.
Our goal is to offer practical & affordable data science education to millions of people around the world, and joining Y Combinator is going to help us execute on our vision faster.
We launched the bootcamp because we saw a need for a more hands-on and intensive program to help our users make a career transition to data science. However, we will continue to offer more free courses and improve the notebook-based learning platform.
In all seriousness though, VC funding is mainly a function of market size and scalability of the business model.
There's currently a huge shortage of data science talent across the world, and a sustainable business funded by venture capital can help fill this gap quickly.
We offer several free courses, and at the end of every free course, students build projects. They vary greatly in scope and size. The 160,000 number refers to the total number of projects created by our users.
Students who are part of our 6-month program build guided portfolio projects that require at least 30-50 hours of work and must meet our evaluation criteria.
I see your point though, perhaps we should use different terminology for the two types of projects.
Congrats on the launch and great progress you've made so far! Setting up a warehouse must be a lot of work, how do you plan to scale this to multiple cities and countries?
In many cases, the existing product is not designed to serve the needs of all countries/demographies/races/genders etc., so there's value in provided a better user experience for a specific niche, especially one that's growing.
I like using products that were created keeping me in mind as the primary customer. Wouldn't you too?
Thanks for your support! We also offer several free courses, so do check them out if you're curious: https://www.jovian.ai
Our program is designed primarily for working professionals with some background in programming, analytics or statistics. We've found that most of them are comfortable with paying program fees upfront or monthly.
However, we do also offer an ISA option in cases where students are unable to pay upfront.
Hi HN, we're Aakash and Siddhant of Jovian (https://www.jovian.ai/). We run a 6-month program that helps working professionals with a background in programming, statistics or analytics land their first data science job.
Four years ago, we quit our jobs and spent several months learning data science and deep learning. We completed several online courses, read some books, and participated in some Kaggle competitions. We were finally able to make a career transition to data science, but looking back at the journey, we felt “It shouldn’t be this hard!”
While there are various platforms offering data science education, most learners don't feel confident while approaching real-world projects even after completing multiple courses and earning certifications. They find it hard to clear interviews to land their first job, and often require several months of training on the job.
Jovian is the platform we wished we had access to when we started learning. There are three things we focus on: (1) we teach practical skills that students use to build real-world projects, (2) we offer 24x7 support & detailed feedback on every project, (3) students learn together with a community of like-minded learners in a university-like environment. We’ve taken inspiration from courses like fast.ai and mlcourse.ai.
Our 6-month program consists of 7 courses, 14 coding assignments and 4 portfolio projects. Every week, students watch 5-6 hrs video lectures and work on assignments for 15-20 hrs. Students can ask questions and get help 24x7 from our TAs over Slack and Zoom. Every 5 weeks, they complete a unique guided real-world project and publish a blog post. We offer 1:1 resume review, interview practice, and job search support for 12 months. We make money by charging a flat fee for the course, which students can pay up-front or monthly.
One thing we discovered and found surprising is that most of the math used in data science can be explained through code. Many people find math intimidating not because the concepts themselves are tough, but because they aren't used to mathematical symbols and equations.
TL;DR: Keep it flat, obvious and simple. 7 top-level folders (app, components, database, emails, functions, lib, public) instead of deep nesting. Explicit server/client separation. Everything has one obvious place to live.
Happy to answer questions about working with LLMs on larger codebases or specific patterns I've found useful.