github.com/b44ken linkedin.com/in/borattob
i go to uoft. i like building stuff
finetuned qwen 14b to talk like me. downloaded all my discord chats via a tool, formed a dataset of 250ish samples and used unsloth to finetune qwen. then i built a chrome extension to feed input events to discord (in the future i may support other apps). it runs locally (ollama). i did something similar when gpt3.5 fine tuning dropped but honestly this works better
most ai study apps are gpt wrappers that don't really teach you, because the core of a chatbot is just answering questions. but the core of a tutor is that it genuinely tests you, then teaches. here, you upload homework, it breaks down what you need to know, and then loops practice question (show work) - long form feedback - retry til success
this is better than gemini quizzes, say, is because the core is long form questions on your material. contrast with a side feature you must seek out, that just asks multiple choice, about what you believe is needed. most importantly, with reactutor, if you cant do the questions, you'll know very well what specifically you need to keep studying
so i built that (powered by gemini 3 and vercel and stuff)
a nice logic gate simulator. i would absolutely love to see this become a 'how to build a computer' tutorial/visualization but that's a long ways out
started working on a wasm4* runtime that i plan to run on a pi pico. right now it implements the framebuffer itself, input, and some calls. this is my first not-small project in rust btw
*wasm4 is a 'virtual console' where games are distributed as webassembly binaries. the binaries expose memory containing a 160x160 display and some other stuff like keyboard input, which you write to and read from to play the games.
built a moduar dashboard (dash.boratto.ca) with the intent of leaving it up on my tv. it displays my current spotify listening and ttc next bus times from their respective apis
built a nice course scheduler for uoft (pulling directly from ttb.utoronto.ca). later i'm planning on allowing users to upload screenshots of course options & parsing them. this way, someone from any school can use it (imo this is a big gap, manually typing in course stuff is soooo tedious)
a toy machine learning framework i built to learn (sequential) neural networks. You can do things like model = Model({ Linear(1, 1) }); to declare a model. See the repo's examples for, e.x. an MNIST classifier.
at winhacks 2025, i delivered a workshop titled write doom in an hour: i and the participants built a rudimentary rendering engine (a raycaster) and a game on top of it. The core working principle is that we divide the screen into a number of top-to-bottom 'strips', and the height of the wall for a given strip is purely a function of its distance to the player. technically this is more wolfenstein than doom, but…