Atharva Mhaske

Atharva MhaskeSep 1, 2026

And just like that, GSoC comes to an end. ❤️ Really grateful to Open Science Labs for giving me the opportunity to spend this summer contributing to open source and learning along the way. Big thanks to my mentors, Ivan Ogasawara and Sai Sanjay Kottakota for all the guidance and support throughout the program. I’m really happy I got to experience GSoC and everything that came with it. Definitely going to be one of the summers I remember for a long time. Now I’m looking for what’s next. 🚀 If you’re building something interesting, solving hard problems, or know someone who is, I’d love to join your crew. Thanks Atharva :)

Atharva MhaskeAug 20, 2026

few weeks back, i learned what instrumentation actually is while working on a project, which led me to explore the otel-go-instrumentation repo. making good, non-sloppy open source contributions is still a thing. this took a lil while, but tbh i learned a lot of new stuff along the way. now i can also add govulncheck scanning to the ci of my personal projects, so that’s a nice bonus. genuine contributions and supporting the projects you actually use will always be better than contributing just for the sake of an open source program.

Atharva MhaskeAug 13, 2026

long time since i posted here so i recently built an ai homelab, just to get hands dirty on gpus with help of kubernetes. for this specific reason i chose quantized model as you know im gpu poor guy jk. my only aim to see how good are these models when quantized and how day2 ops look like for llm workloads on k8s. the stack is simple: i own nvidia l4 gpu(gets job done as of now) on jarvislabs(vendor) with: docker, k8s, kind, nvidia container toolkit and device plugin, bifrost ai gateway and ofc their cli as well, openwebui, kube-prometheus-stack, nvidia dcgm operator, keda for autoscaling(although here max:1) i have opted a fallback support for my provider support like i use cursor alot so it fallbacks as default and vm stops. although here i have opted aggregated setup(will add prefill-decode disaggregations once i get handy with this) Maxim AI bifrost is really amazing, its blazingly fast and consumes very less resources in comparison with other gateways and they also have in house guardrails and monitoring support, which i haven't tried yet but kept it for next versions. i chose Open WebUI for sanity check for local quick consuming and debugging. tbh i learned alot how this all operators go hand in hand to serve models using k8s, one thing is very clear k8s i now not only a control plane for web apps but also for any ai workloads, this is still a naive i haven't optimized every ounce of it as im still exploring. if you are working with llm inferencing in production i would happy to learn from your lessons. also any cloud providers kind enough to give free gpu credits?