Gurunath

AI Inference Engineering

[Gurunath @rajaguru95]

Thank you so much for the shoutout @_nilsbraun .

Doing open source work was addicting/rewarding due to humble/great folks like Nils Braun! ♥️.

"Open source contribution is new competitive programing " - Change my mind 😅

#OpenSource #dask #Python #leetcode

[Nils Braun @_nilsbraun]

The first release of dask-sql as a @dask_dev-contrib member is out! Special thanks to rajagurunath for implementing all the cool new ML features and thanks to the @dask_dev community for accepting dask-sql as a contrib package!

[Gurunath @rajaguru95]

Blog post: https://t.co/yspD3s6XsE

Intuition: Typical LLM inference workloads follow this pattern: busy hours stay busy, and idle minutes are followed by idle minutes. The lag correlation is high. Reference (Year serving LLM inference workloads): https://t.co/TfSx7k6d3u

[Gurunath @rajaguru95]

https://t.co/2Hh79yvj7h

[Gurunath @rajaguru95]

Introducing Tidal - co-serving online + batch LLM workloads on the same GPU.

The idea: use GPU capacity left idle by online traffic for batch work, while protecting online latency.
Paper: https://t.co/el0fCK1tsv
Project: https://t.co/wtwADLlPkZ #LLMInference #vLLM #GPU #sglang

[Gurunath @rajaguru95]

https://t.co/2Hh79yvj7h

[Gurunath @rajaguru95]

@ionet - Generated with MiniMax H3 on the https://t.co/z5bpLCNcVt CaaS recipe.
- Video and audio in one pass, one GPU. 25 seconds, 1440p.

[Gurunath @rajaguru95]

@jpschroeder Qwen 3.6 27B(dense), which is priced higher than 35B (moe) even though the demand for 35B is higher. But 27b performs well with artificial analysis benchmarks ..

So it's not based on model demand only right?

[Gurunath @rajaguru95]

@Kimi_Moonshot Makes sense :) 🙃😁

[Gurunath @rajaguru95]

@emergentlabs @buildwingman @nitprashant Architecture + all 3 Skill files shared below
Github: https://t.co/p3Wh1U9z9p

Built on @buildwingman by @emergentlabs

[Gurunath @rajaguru95]

"Every company needs an OpenClaw strategy."
— Jensen Huang, GTC 2026

So far, that strategy has only lived in engineering.
We built it for everything else.
Day 2 · #Wingathon 🧵 @emergentlabs @buildwingman
@nitprashant https://t.co/SGHYjLRpAY

[Gurunath @rajaguru95]

@emergentlabs @buildwingman @nitprashant The goal:
Let a solo founder start from the second
floor (launchpad). Not from scratch. Not alone.

With Ayan running marketing, Kiyan watching the money, and Ziyan keeping customers happy.

You start where a funded 5-person team would.

[Gurunath @rajaguru95]

@emergentlabs @buildwingman @nitprashant Introducing StartupOS - 3 Wingman agents
for the non-technical half of your startup:
🟣 Ayan — Marketing Claw, Ads, content, Remotion videos
🔵 Kiyan — FinOps Claw, MRR, churn scoring, infra cost,
🟢 Ziyan — Customer Retention Claw
Help articles, MCPs, ticket intel

[Gurunath @rajaguru95]

@beknabdik @buildwingman @emergentlabs 100% fair. Word-level from ElevenLabs --json is our v1.

Enough to catch dropped syllables and stress, but not partial phoneme errors.
Forced alignment is next. Thanks for the sharp take. Lots to improve our hackathon solution, I guess.. : )

[Gurunath @rajaguru95]

We shipped Saras 🪷 on Wingman for #Wingathon. Duolingo is a public app. Saras is your private agent, rips your anime/movies into lessons, and sharpens your accent by the day.
Github: https://t.co/BWrmh3wfWD
Video: https://t.co/96l22rajI5

On @buildwingman by @emergentlabs

[Gurunath @rajaguru95]

@buildwingman @emergentlabs Why Saras?

🧠 Every chat feeds your personal KB (karpathy-wiki). It gets smarter about YOU with each chat.

🎬 Lessons ripped from your anime & movies.

🛠 Built by you, for you.

[Gurunath @rajaguru95]

@buildwingman @emergentlabs All of this is ONE SKILL.md on Wingman. No mobile app, no backend API, no cron server — Wingman's runtime does the plumbing.
Full architecture + single-block diagram: https://t.co/2nH5kkYhRU

Thanks @buildwingman @emergentlabs for the primitive 🫡

[Gurunath @rajaguru95]

@buildwingman @emergentlabs Anime-style and movie-style videos generated by Saras-wingman
Built with @saran_gogosoon @AI_Techie_Arun

[Gurunath @rajaguru95]

@TheAhmadOsman @grok two questions on this article
1. How the concurrency fits in this equation !? How to calculate minimum- maximum concurrency possible with the given model weights , VRAM , quantization etc ?

2. What is the similar forumula to calculate KVcache memory utilisation?

[King @0xChesterKing]

Shoutout to @rajaguru95 for using @novelecosystem to build a seamless treasury agent. More agents coming soon! 🚀 If you're building with IO intelligence built by @ionet and want to integrate blockchain operations, my DMs are open.

[Gurunath @rajaguru95]

Happy to share my new blog post on Trino SPI plugins! If you're a Pokemon Fan, you're gonna love it! And all you @trinodb and LLM enthusiasts out there, it's a great read too! Check it out 👉 https://t.co/9AAcoobVsq

#OpenAI #Java
#Trino #Pokemon #LLM #DataEngineering"