After CompSci

On weekends and early mornings, I try to work with new tech as it comes out. This past week, it was PRDs for genetic programming and Railway for hosting the apps.

The project is relatively simple and has been on my Google Keep list for at least a year. A calculator that takes in the model, the number of tokens each user uses, and the model you want to charge for using those tokens. The result is the minimum amount you should charge for your service. The goal is to give makers a clearer idea of what they charge for the base cost of their app when using LLMs or other back-end models. A lot of the time, indie hackers and engineers aren’t the best at setting prices, so this at least gives them a hand at basic finops.

I started out by writing up a PRD.md file. It included the idea of the following:

• An outline of the project

• The format of how the project should be structured

• The formula for the calculator to use

• What tools and languages should be used for the code
• How to make the deliverable, as Docker images

To me, this felt a little more like one-shotting a project, but in a more refined format. The initial product worked pretty well. I’ll be using this method for future projects. Troubleshooting UI bugs was difficult, though. Frontend development is an area that I am less entrenched in, so I think part of the issue is me. I’m sure I’ll pick up proper CSS one day, though. Coding assistants are feeling more robust now, though, as they can even help point out some flaws. When I had asked for a UI fix, the fix was there, but I had not rebuilt the containers before testing locally, a stupid human error, but one we have all encountered. Codex asked me to confirm I was rebuilding, and what I saw was the UI working as I asked it to.

The deployment process and the use of Railway were relatively simple. I had heard good things about it from some buddies and decided to give it a try. I’ve been building dependent workflows for my day job for almost a decade now, and Railway reminds me of some of the Heroku tooling. I liked the ability for it to both deploy from the cli and also from commits to the repo. I still need to create a proper railway config, but clicks were enough for this weekend’s project. Linking up a repo was simple, and deployments were quick. The UI for searching logs was pretty solid, and the alerts for failed deploys out of the box are a nice feature.

If you are looking to price out your services and get an idea of your cost, try out the AI Token Pricing Calculator.

Final Notes from this week’s learning:

• Railway is great for deploying small projects quickly
• PRD driven development can help you get through your project list and is definitely a form of what is here to stay with agentic workflows
• Having consistent deployment patterns that allow you to test changes early and often will be even more important as we more agenticly built software
• The tooling is moving at a pace unlike anything I have seen, so take small bites each week to test out, so you can stay in the flow of where industry is going without feeling overwhelmed.
• The agentic coding models are improving at some platform and infrastructure tasks. I was able to get Dockerfiles made for each service with a single prompt.