SL Available for work

Product engineer / Obvix Labs researcher / Delhi

Hard problems.
Working software.

I turn rough ideas into dependable products. I build across the stack, use AI as a serious engineering tool, and stay with the work through the unglamorous last mile: tests, failures, deployment, and real users.

See selected work Tell me what you need
BUILD LOG / 09.26 SYSTEM READY

01

From blank file
to a live URL.

FRONTENDshippedBACKENDhardenedDEPLOYlive
Production
Products live in the wild
Verification
73 backend checks on Fridgey
Research
Researcher at Obvix Labs
Timezone
UTC +5:30 · remote worldwide

01 / Selected work

Products with
proof behind them.

Live software, technical depth, and the decisions that made each build hold up.

Live product · 2026

Fridgey

Turn what is already in your kitchen into a practical meal plan.

I built the product surface, authentication, data layer, AI pipeline, quotas, subscriptions, and deployment. The interesting work lives after the happy path: validating model output, recovering from failures, guarding duplicates, and fixing a PWA cache bug that served new HTML with stale JavaScript.

  • Gemini + Supabase
  • Netlify Functions
  • Lemon Squeezy webhooks
  • 73 backend checks + 3 browser suites
Open live product
LIVE / PRODUCT
9:41● ● ▰
Ffridgey•••

YOUR KITCHEN

What can we
make tonight?

🥬 Spinach🍅 Tomatoes🥚 Eggs＋ 6 more
⌂◎□
MON / 24 MINRoasted tomato
shakshuka680 kcal
TUE / 18 MINGreen pantry
pasta540 kcal

Live consumer product

heartcoded.store

A custom-page product with Razorpay payment verification, Cloudinary uploads, and an automated fulfilment flow. Designed, built, deployed, and operated end to end.

Next.js 16RazorpayCloudinary
Visit the product

02 / Research & systems

I do not stop at
“it seems to work.”

Research trained me to separate a good-looking result from a defensible one. That instinct carries into every product I ship.

Economics · Under peer review

Two faces of
algorithmic collusion

Memory, spuriousness, and a model-free entropy diagnostic in homogeneous-good Bertrand competition.

Two pricing algorithms can settle on high prices without communicating. My paper shows why memory changes whether that behaviour is durable collusion or a misleading surface pattern—and proposes a diagnostic to tell the two apart.

Written, coded, and submitted independently.

Read the plain-English version
StatefulStateless
high-price regiontraining →

The same outcome can hide a different mechanism.

A / OPEN SOURCE

EconGym

Gymnasium-style economics environments with shared interfaces, equilibrium benchmarks, and tests—built so standard reinforcement-learning agents can enter a market model without custom glue.

View repository

B / SECURITY TOOLING

AgentReady

An auditor for whether commerce sites are legible to AI shopping agents. It renders pages in Chrome, reports only high-confidence defects, and closes SSRF at both fetch and browser layers.

Technical write-up available on request

C / PRODUCT AUDIT

19 defects, reproduced

A commissioned teardown of a shipped language-model feature, including prompt handling, output validation, failure behaviour, and eight adversarial injection paths.

Client-confidential engagement

03 / How I work

AI in the loop.
Judgment stays mine.

I work inside Codex, Claude, and other agentic tools every day. The speed is useful. The responsibility cannot be delegated.

  1. 01

    Find the real constraint.

    Define what must work, what can wait, and what evidence will count as done.

  2. 02

    Build the narrowest complete path.

    Ship one end-to-end workflow before multiplying features or abstractions.

  3. 03

    Attack my own assumptions.

    Read the generated code, test the ugly paths, instrument failure, and verify external boundaries.

  4. 04

    Stay after launch.

    Watch what breaks for someone who did not build it, fix the friction, and leave the system easier to own.

A short introduction

I am Shubh, a self-taught product builder and researcher at Obvix Labs, based in Delhi. My path is unconventional; my work is public, testable, and easy to inspect.

I am strongest where product judgment and implementation meet: taking an unclear idea, finding the useful core, building it, and getting it onto the internet.

I prefer async work with clear outcomes, direct written communication, and enough ownership to solve the actual problem.

Build with
JavaScript, TypeScript, Python, SQL
Systems
Supabase, Netlify, APIs, payments, LLMs
Languages
English, Hindi
Links
GitHub ↗ Instagram ↗

04 / Contact

Have a real problem?
Put it here.

Tell me what you are trying to ship, what is stuck, and what a useful first result looks like. I read every message myself.

Prefer email?shubhhdoespy@gmail.com

Encrypted in transit and saved to my private inbox.