I want to become a full-stack AI engineer, and I'm working toward it. This site is one of the things I'm building on the way.
About
C · C++ · Python · Go · TypeScript
01 · Main project · Python, Windows
Natural voices for every Windows app. Facts: only apps that use Windows voices
I was reading ebooks and the built-in Windows voices were painful to listen to. Good AI voices already existed, but my apps couldn't use them. So I built the bridge.
Facts: checked on GitHub
The old bookshop on the corner had a peculiar charm about it. Dust motes danced in the sunlight that streamed through tall windows, and the smell of aged paper filled every room. It was the kind of place where you could lose an entire afternoon without even noticing.
Switch voices while it plays. It picks up at the same sentence.
An illustration. AI voices sit in the normal Windows list, so the app doesn't need to change.
02 · Building now · TypeScript, a local model
A Discord bot for a server whose friends speak different languages. Everyone writes in their own language and reads everyone else in theirs, under each person's real name.
The translating is done by TranslateGemma, a 4B model running on the server owner's own PC or Mac, so there's no API bill and the messages never leave that machine.
Example translations written out ahead of time, so you can see the flow without a model here.
Smaller builds
It reminds me about the LeetCode daily, reads my wrong submissions, and hints at what went wrong. The hints get stronger through the day, and it can auto-solve at most twice a week so it never becomes a crutch.
Example messages for a Hard question.
You talk to Blender. A model writes the Python and runs it, and when it fails, the model reads the error and tries again. Simple commands work; it doesn't model anything you can say.
A replay of how the loop works, not a live model.
A small Unix shell: it parses a line, runs its five built-ins itself, and finds anything else on the PATH to fork and exec.
The built-ins, simulated from the README.
A reverse proxy for HTTP/3 in Go, built to learn how the protocol works below the API. Stop a backend and it leaves rotation after one failed request; start it again and it's back after two health checks, about 6 seconds.
0 served, 0 failed. The health check timing is real. A simulation of the behaviour.
Handles 1000+ connections. Facts: not measured here
One C program that serves web pages and chat. Before it can answer anyone, it has to work out what the incoming bytes are. Type some.
The same check as server_detect_protocol in server.c.
A 3D map of India's rivers. First the big picture of where the water goes, then a guided journey down one river, with stops you can step into. A side project, still in development.
A sketch of the journey's eight stops. The real one is a 3D map with 255,319 mapped river sections.
A learned controller picks the compression. Facts: a simple rule beat it
A course project. A coded camera blends several frames into one readout, which saves readouts and blurs road signs. I detect signs on those readouts directly. The camera blends twelve frames while the road is empty, and drops to single frames when it sees a sign.
0 frames, 0 readouts: 0% fewer readouts.
Top row: frames. Bottom row: readouts. A simulation of the rule; on the test videos it used 71.3% fewer readouts.
That's everything so far. I'm learning how agents work by building them, and whatever I ship next lands here.
Each crew builds the same page its own way. If you don't pick, the Studio starts on its own.