Skip to content

Vanshaj
Poonia

I build software across the stack, from product interfaces down to a hobby operating system in Rust, and everything here says what state it is actually in.

Full Stack and AI Developer, Jodhpur, Rajasthan

Email meGitHubRésumé

ChronoOS

A hobby operating system written in Rust.

Stack
Rust, no_std, x86_64
Verified
Boot to shell, typed input, a ChronoFS session
Deployment
Local only, runs locally in QEMU

Active

Case study

Calibrated 2026-11-12

Selected work

Three projects that show the range. The rest are on the work page, with the state each one is in.

All work
ChronoOS boot path

  1. Firmware / QEMU

    loads the bootable image

  2. Bootloader

    long mode, paging, framebuffer + memory map

  3. Kernel entry

    no_std Rust, serial before pixels

  4. Interrupts + memory

    GDT, IDT, PIC, PIT, free-list heap

  5. ChronoFS

    ATA PIO, fsck, repair, journal

  6. Shell + apps

    commands, eras, notes, calc, sysinfo

ChronoOSrepository: chronosapien

A no_std x86_64 kernel that boots under QEMU, draws its own framebuffer console, and hands you a shell. Built to understand boot, memory, interrupts, and filesystems by writing them.

State
Active
Deployment
Local onlyRuns locally in QEMU
Stack
Rust, no_std, x86_64, QEMU
Case studySource
ATLAS request path

  1. UI

    shared shell, palette, boards, calendar

  2. Capability layer

    one place that answers can-this-user

  3. Supabase session client

    no service-role bypass, ever

  4. Postgres + RLS

    the boundary the AI obeys too

  5. AI tools

    workspace reads, web search, URL fetch

Project Management ATLAS

Boards, custom fields, saved views, a marketing content calendar, and an assistant that reads the user's real workspace through the same row-level-security boundary the UI uses. Deployed for one organisation.

State
Active
Deployment
PrivateDeployed behind sign-in for a single organisation
Stack
Next.js, TypeScript, Supabase, Postgres RLS
Case studySource
TensorTrail gradient flow

  1. Tensor

    NumPy array that remembers its origin

  2. Operation graph

    built at runtime, node per op

  3. Topological sort

    reverse order for the backward pass

  4. Autograd

    local derivative per node

  5. Broadcast reduction

    gradients summed back to shape

  6. Optimiser step

    SGD, momentum, or Adam

TensorTrail

Reverse-mode automatic differentiation, broadcasting-aware gradients, convolutions, optimisers, and a trainer, written so the machinery stays readable. 137 tests and a finite-difference gradient checker keep it honest.

State
Shipped
Deployment
Local onlyPython package, installed locally
Stack
Python, NumPy, pytest, GitHub Actions
Case studySource

Open source

6 pull requests merged and 23 open, across 9 projects I did not start. Each status was checked against GitHub on 2026-11-12.

All contributions

I like moving between abstraction layers.

The range is not scattered. Each layer down exists because the one above it raised a question I could not answer.

  1. Product

    Applications people sign into and use to do their job.

  2. AI systems

    Routing, tool use, memory, and metering around hosted models.

  3. ML internals

    The layer under the API: autograd, gradients, optimisers.

  4. Developer tools

    Editors, runners, and the infrastructure that hosts them.

  5. Systems

    Boot, memory, interrupts, filesystems. No runtime to fall back on.

Experience

Full experience

  1. Helix Grid

    Software Engineer

    Now

  2. Open source

    Open Source Contributor

    Now

  3. codedamn

    Campus Ambassador

    Feb 2022 - Aug 2022

  4. Symbl.ai

    Open Source Developer

    Oct 2021 - Aug 2022

Workbench

Build logs, debugging write-ups and design studies. Rough on purpose.

Open the workbench

One study, still running.

I am working on a reproducible study of training-time masking: whether deliberately hiding features during training makes classical tabular classifiers hold up when the data they meet later has gaps the training data did not.

In progress, unpublished

Research

Have an interesting problem?

I am most useful on the ones that cross a boundary: a product decision that turns out to be a data-model decision, or a bug that only reproduces in production. If that sounds like yours, I would like to hear about it.

Email meGitHubLinkedInRésumé

vanshajpoonia06@gmail.com