Skip to content
View sarthaknimbalkar's full-sized avatar

Organizations

@Vervelio-Labs

Block or report sarthaknimbalkar

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
sarthaknimbalkar/README.md
Sarthak Nimbalkar, independent software engineer building autonomous AI systems. A year of contributions burning quietly in the dark.

Sarthak Nimbalkar

Independent software engineer building autonomous AI systems end to end, from the data layer to the model to the interface. Works remotely, shipping since 2019, and co-founder of Vervelio.

Website · Vervelio · LinkedIn


Every server you own just went unreachable in the same second.
You are the one on call. Ten minutes of your life.

Pager alert at 2:14am: every region, every instance unreachable in the same second.

38 states · 322 ways through · 7 endings, one of them correct · fastest the game allows 19 min · my time, the night it was real, 7 min

Warning

Most people restart the fleet. It works, for eleven minutes.

It runs inside GitHub with no JavaScript, because GitHub does not permit any: every state is a pre-rendered file, every move is a hyperlink, and a theory you have disproved stops being a link at all. Same trick as DOOM-in-a-README, pointed at an outage.

The thing that actually caused it, if you would rather read than play
A 412 millisecond dependency fault fails 8,800 clients at the same instant. They all retry at exactly one second, arriving inside one millisecond, which fails them again. The loop feeds itself and saturates the gateway for two hours eleven minutes. Random jitter across a 900 millisecond window drops the peak from 8,800 to about 10.

Six weeks earlier, someone had made the client resilient to blips: six retries, one second apart, no randomness. Clients that fail together retry together, so they fail together. The loop has no exit of its own, and the fix is one line of arithmetic. With N clients retrying on a fixed beat, all of them land inside the same processing window Δ. Spread that beat across a random window J instead:

$$\text{peak}_{\text{fixed}} = N \qquad\qquad \text{peak}_{\text{jitter} \sim U(0,J)} \approx \frac{N\Delta}{J}$$

At N = 8,800, Δ = 1 ms, J = 900 ms, the peak falls from 8,800 to about 10.

Randomness is a safety feature. Anything that fails together will retry together.

The failure mode is a retry storm, and it keeps happening because backoff alone looks like the fix. Backoff changes how often one client retries; it does nothing about eight thousand clients counting in step. The reference is Marc Brooker's Exponential Backoff And Jitter, whose formulas now ship inside every AWS SDK.



What the graph doesn't show

I build systems that run without me. Four are in production right now, with real users and real money. Most of the time, nobody is watching them work. Not even me.

It didn't start there. In 2019, I could barely make a script do one thing. So I built the next thing before I felt ready, shipped it, and let what broke teach me. Then I did it again. Scripts became apps. Apps became the systems under them. Systems became models. Models became software that runs itself. Same move every time, one step higher. That path led to Vervelio, the studio I co-founded to build and ship its own software.

The hard part was never the code. It's the judgment. An agent working alone at 4am is just running a decision I already made once, well enough that it holds while I sleep. A system is a cast of its maker's mind. So I don't build timid ones. A tool nobody dares trust proves nothing about the person who made it. Mine run for fourteen hours with the room empty, and I stand behind every call they make in it.


sarthaknimbalkar.in · what building actually costs, not the version that sounds good afterward
vervelio.com · LinkedIn

Sarthak Nimbalkar · independent software engineer · autonomous AI systems · remote

I am Sarthak Nimbalkar, an independent software engineer and co-founder. I build autonomous AI systems end to end: the data layer, the model, the infrastructure, the interface, and the part that takes money. I work remotely and I have been shipping since 2019.

I build agentic systems that can be trusted with real users and real money, and the command layer that keeps them my instruments instead of my replacement: cost ceilings and spend controls, rate limiting, audit trails, approval gates, and hard boundaries on the decisions that stay mine.

Work agentic AI and multi-agent systems · large language model applications and orchestration · LLM cost control and observability · AI governance and guardrails · human-in-the-loop review systems · retrieval and data pipelines · local-first and self-hosted infrastructure · full-stack product engineering from database to interface · solo end-to-end product development
Stack whichever one the problem deserves. It is 2026; asking me for my stack is like asking a carpenter which hammer brand he is loyal to. Picking one up takes me a weekend. The stack was never the hard part, and anyone still selling theirs as the skill is telling you where their ceiling is.
Currently building products at Vervelio, a studio that builds and launches its own software rather than waiting to be hired
Writing sarthaknimbalkar.in, on what shipping real systems costs
Open to technical co-founder conversations, hard problems in autonomous systems, and work where one person owning the whole stack is the advantage rather than the risk
Reach me LinkedIn
Frequently asked

Who is Sarthak Nimbalkar? Sarthak Nimbalkar is an independent software engineer and the co-founder of Vervelio. He builds autonomous AI systems end to end, from the data layer and the model to the infrastructure and the interface. He works remotely and has been shipping software since 2019.

What does Sarthak Nimbalkar build? He builds agentic AI and multi-agent systems that can be trusted with real users and real money, together with the control layer that keeps them accountable: cost ceilings, rate limiting, audit trails, approval gates, and human-in-the-loop review. Four of his systems are in production today.

What is Sarthak Nimbalkar's tech stack? Whichever the problem deserves. He treats the stack as interchangeable rather than an identity, and picks the language, framework, and infrastructure that fit each system, from local-first and self-hosted setups to full-stack product engineering across the whole database-to-interface path.

Is Sarthak Nimbalkar available for work? He is open to technical co-founder conversations, hard problems in autonomous systems, and work where one person owning the entire stack is an advantage. Reach him on LinkedIn or read his writing at sarthaknimbalkar.in.

Everything on this page is generated by scripts in this repo: the on-call incident and the banner, re-rendered on schedule by Actions. No stats services, no badges, and no JavaScript anywhere, because there cannot be any.

Pinned Loading

  1. anchor anchor Public

    Make your AI coding agent actually follow your rules, every turn - re-injection, verification, and correction that survive context compaction. Local-first, Python, works with Claude Code.

    Python

  2. stack-swap stack-swap Public

    Swap whole GPU project stacks on a single shared box with one command, over SSH. One GPU, multiple projects: flip between them instead of editing configs by hand. Tiny, dependency-free Python.

    Python 1