Atharva

Back to experiments. The goal now isn't getting better numbers, it's understanding why the numbers move in the first place. A lot more ablations ahead. updated readme: github.com/Atharva-Mendhu…

Quick StrataRL update. First Kaggle run and now I know why training the model is the easy part. Getting the infrastructure, logging, monitoring, and evaluation right took way more time than I expected. Starting to understand why people obsess over metrics so much.

@iyoushetwt Just because of the hardware this might be true

unpopular opinion: macOS is better than linux (for coding)

@itzsam_ai Im solo dev so idk 😝💔

@atharvarta when will you launch this?

Are you Building in public? Drop your project below👇

@nezbuilds Working on StrataRL: A GRPO infrastructure for multi-domain reasoning in Small Language Models. github.com/Atharva-Mendhu…

Good morning builders 👋 It’s Wednesday, time to show the internet what you’ve been building. Drop your project + a short description below. I’ll be checking out projects, giving feedback and connecting with fellow founders throughout the day 👇

Local validation is done. Now I'm moving the experiments to Kaggle The biggest bottlenecks so far are GPU memory constraints, rollout speed, and fitting meaningful GRPO experiments into Kaggle's runtime limits Would love to hear from anyone who's run RLHF/GRPO on Kaggle before

GRPO + domain-aware rewards + stratified advantage normalization + curriculum scheduling + training monitoring Currently validated locally on a Laptop (MBA M4 24GB) with Qwen2.5-3B, 35+ tests passing, 0 failures for now, and successful training runs so far

Lately I've been reading a lot about RL and GRPO. What started as a few papers turned into a deep dive on why multi-domain RL can improve one benchmark while making another worse. One paper I found interesting was DeepSeekMath. Worth a read if you're interested in GRPO and RL.

Just finished reading SICP Chapter 3. Never thought a book written decades ago would map so well to modern agentic AI systems. State, shared context, mutation, concurrency, synchronization. Same problems, different scale.

@buildwithsid github.com/Atharva-Mendhu…

kindly share your github profile i wanna judge you

@yahyavision mendhu.tech

Drop your portfolio URL Let’s drive some traffic

Started SICP thinking it was a programming book. It’s more about reasoning (which I like) than coding (which I dont like) i like this book

@KaiXCreator Windows has the biggest MOAT

be honest, if Linux is so good why do most people still use Windows?

@cneuralnetwork working for this happen

weekend

@YashHustle_22 Linux macos solely for the macbook

Which operating system is worth using for coding? - Windows - MacOS - Linux

@KaiXCreator You cant LARP with Mac mini to cafés

As a developer, which one would you choose? Mac mini or MacBook

@shub0414 bloat

Tell me Linux distro with 0 haters: I’ll go first

@programmerByDay Productivity & Collaboration floework.vercel.app

Your SaaS. Only 3 words. Let’s hear it 👇🔥