How did I get into @HackTheNorth, after 3 years of rejections. 😭. If anyone wants to team up, then let me know!

Had a quick little side-quest at NASA 🚀

I was in Toronto, and this happened.

Took 2 all-nighters to build an MCP (Model Context Protocol) powered car. Don't regret a single second of it.

Guess who won 2nd overall at Hack The 6ix 🥳

Here is a chopped video of me explaining the app I've been working on for the past month, Rapix . p.s. I need only 6 more Android beta testers before I can upload it to the PlayStore. Reply if interested 🙏

Side Note: I need 12 beta testers before I can launch the app. Feel free to DM me if you'd be open to it. It's barely any work :).

Does 'Prabh Studio' go hard? 👀

We just won Snowball iCE microphones @ Canada's largest high school hackathon (@jam_hacks) 🥳. Welp, thanks for the birthday gift JamHacks :)

Hyped for @jam_hacks at @UWaterloo 💪🔥. cc: @evelynhannah_ @virkvarjun

I'm at @UWaterloo with @virkvarjun from the 16th to the 18th. Hmu if you wanna meet up.

8/ Curious? I wrote an in-depth article on this with mathematical explanation's. I'm building more AI projects like this, follow me for more :) . And if you’re ready to tinker, the GitHub repo linked above has everything you need.

7/ Why bother?. Compression opens doors everywhere: object-detection in autonomous drones, diagnostic cues on portable medical scanners, battery-friendly IoT cameras that only wake the network when something matters, and the list goes on. Compact transformers = powerful,.

6/ INT8 quantization:. Weights and activations are mapped from 32-bit floats to 8-bit integers using per-channel scale and zero-point values learned during a brief calibration pass. This cuts model size 4x and speeds up integer arithmetic on CPUs, while keeping performance

5/ Structured pruning:. After training, we measure how much each attention head and MLP unit contributes. Units with the lowest L₂-norm scores are removed in whole chunks (about 30% of heads, 25% of MLP channels). A short fine-tune lets the remaining units adjust, giving us a

4/ Knowledge distillation: . Imagine a pro photographer teaching a beginner. Instead of only saying “That’s a cat,” the pro also explains how sure they are and what features hint at “maybe a fox.” . A smaller ViT is trained to match both the class probabilities and the

3/ I was able to reduce the size of the model without greatly affecting accuracy using three techniques: . - Knowledge distillation (learn from a bigger model).- Structured pruning (delete lazy attention heads & MLP units).- INT8 quantization (store thoughts in 8-bit integers).

2/ Therefore, I built Pocket-ViT. A Vision Transformer slimmed from DeiT-Small’s 22 M parameters down to ~5 M, then squashed to a 1.3 MB INT8 file, small enough to ship inside a mobile app update.

1/ Most of our cameras are on our phones, drones, or Raspberry Pi boards. To use make use of transformer-level accuracy there, we need to compress these models significantly to work effectively on 'pocket size' hardware.

Big Vision Transformers (ViT's) CRUSH accuracy benchmarks, but they also crush your RAM, battery, and download quota. We need smaller ViT's for edge devices, think drones, phones, even Raspberry Pi. My goal: keep the ViT accuracy, reduce the size. Starting from DeiT-Small (≈