What it is
MidNightAI is the part around the model. Inference is a single API call. Everything that makes it usable is application code: knowing who you are, remembering what you asked an hour ago, turning a wall of Markdown into something you can read. That's the part I built.
It runs on infrastructure I control rather than someone else's account, which is the whole reason it exists.
How it fits together
Fast enough to think alongside
Requests go to Groq's LPU inference, which returns tokens fast enough that the interface stays conversational instead of making you wait for a paragraph to finish. The speed is as much a product decision as a technical one.
Accounts with a second factor
Authentication runs on Laravel Fortify with two-factor enabled. A chat history is a record of what someone was working on and thinking about, so it gets treated as sensitive rather than as convenience data.
History that persists
Conversations are stored in MySQL and reloaded with their context, so a thread survives a closed tab. Sessions are handled so that recalling a long conversation doesn't mean replaying all of it every time.
Output you can read
Answers render as proper Markdown. Tables that stay tables, maths that stays legible, syntax-highlighted code blocks. An answer full of raw asterisks is only half delivered.
Deployment
Push, hook, migrate, live
The project lives on a Plesk environment and deploys from GitHub over a
webhook. Push to main, the hook fires, migrations run and
assets compile on the server. Updates go out without downtime and without
anyone copying files over SSH, which matters more than it sounds like it
does when the thing is running and you want to fix something small.
Available for work
Building something on top of a model?
The interesting part is rarely the API call. If you need the application around it, the auth, the persistence, the deployment, an interface people will actually use, that's the work I do.