Divya Kulshreshtha
AI Lab · 2026
Scroll to enter the lab
Home → AI Lab
AI Lab I

Design × intelligence.

Tools and automations I build to take the busywork out of design and research. Some run in production, some are weekend experiments — all of them earn their keep by saving real time.

8 Active Projects
5 Live in production
2026 Ongoing
Waking the Machines
i Field notes

I kept running into questions about AI and design that I couldn't answer by reading about them. So I built things. These are five — some for work, some for myself, most still running.

Ch. 01Design in the real material

I stopped drawing pictures of software. I build it now.

For a long time, design meant making a convincing picture and hoping it survived handoff. Now I just build the thing. This whole site is the example — made in code, not Figma.

Screen recording of the portfolio site being built in code
01 Live

Interactive Portfolio in Code

The site you are reading. Built using Claude Code with an iterative editorial design workflow — brief in chat, deploy in minutes. Demonstrates AI-augmented design practice end to end.

Claude Code HTML / CSS Open Design
Personal 2026
Ch. 02Scaling systems

The handoff was always where things broke.

A design system in a Figma file is something you have to maintain in two places. I built Hiremate's system in code — tokens, components, the lot — so there was only one place. No translating, no drift.

02 Live

Design System in Code

Built the entire design system for Hiremate in code — tokens, components, patterns — while delivering the product itself. No Figma handoff lag. Devs pulled tokens in minutes, built in two days.

Claude Code CSS Tokens HTML / JS
Naukri · Info Edge 2026
Ch. 03New ways to conceptualize

Fast tools make it easy to skip the thinking.

When you can build anything quickly, the temptation is to just start. I made a canvas that forces the problem open before any building begins — what am I actually solving, for whom, and why. The output is a brief an AI can genuinely work from.

03 Live

AI Design Thinking Canvas

With all product thinking and exploration going into AI and code, I built this canvas as a nudge to product thinkers to complete every dimension of problem discovery — then converge it all into a design.md an LLM can consume and build from holistically. Stops the "narrow prompt, shallow answer" trap before it starts.

Claude design.md Product Thinking LLM
Personal 2026
Ch. 04Scaling judgment, not headcount

Research dies in transcripts nobody reads.

Most synthesis never happens. Notes pile up, patterns stay invisible, decisions get made from memory. I built a pipeline that tags interviews as they come in and surfaces what keeps recurring — without anyone needing to find the time.

04 Live

Research Synthesis Pipeline

Fed structured interview transcripts and domain knowledge into NotebookLM to identify patterns and generate RAG-grounded insights. Research stopped decaying in folders and became a living, queryable repo the team could build on.

NotebookLM Claude RAG
Info Edge · Internal 2024
Ch. 05When building gets cheap

When building gets cheap, you start making things just for yourself.

I built a finance tracker for my partner and me — not because no app exists, but because none of them think the way we do. That's new. Most software is made for a market. Now you can just make it for two people.

05 Live

Personal Finance Tracker in Code

Took my code exploration into personal finance — expenses, earnings, investments, and health scoring in one place. Helped me and my earning partner collaborate better and see our financial picture clearly. All designed and built in code.

Claude Code HTML / CSS / JS Data Viz
Personal 2026
ii On the bench

AI didn't just speed up what I already did — it opened modalities I couldn't cheaply touch before. Here's what's live on my bench right now. Recordings landing soon.

Voice

Voice as a design surface

Prototyping interfaces you talk to, not tap. Exploring where voice genuinely beats a screen — and being honest about where it really doesn't.

Microinteractions

Microinteractions, generated

Sweating the small motion. Spinning a dozen variants of a single hover or transition in seconds, then keeping the one that feels alive.

New patterns

Interaction styles, prototyped live

Gestures, generative canvases, direct manipulation — testing interaction patterns that never had a cheap way to prototype until now.

iii Also running

Quieter automations keeping the studio moving in the background.

10 In Progress

Apply Agent

Job hunting has too much copy-paste and not enough signal. This agent crawls the careers pages of companies I actually want to work at, spots relevant openings, drafts a contextual note for each, and queues LinkedIn outreach to the right people. Keeps my profile copy fresh as my work evolves.

Claude API Playwright LinkedIn n8n
Personal 2026
11 In Progress

Figma Legacy to Code

Naukri's design practice grew up in Figma. Moving it to code isn't just a file migration — it's rebuilding how designers explore. This lifts the legacy component library into a code-native system so designers can prototype directly in the medium that ships, without the round-trip.

Claude Code Figma API CSS Tokens HTML / JS
Info Edge 2026
12 In Progress

Version Management for Design Concepts

The hardest part of designing in code is running multiple ideas at once. You either commit too early or lose work. This is a lightweight branching system that lives inside the project — lets you park an idea, switch to another, come back, and compare without needing git expertise.

Claude Code Git HTML / CSS / JS
Personal 2026
13 In Progress

Agentic Email Inbox

Gmail hasn't changed how it thinks about email since 2004. This is a redesign that treats the inbox as a surface for action — generative reply drafting, intent detection, threaded context that surfaces what matters. Built as a working prototype, not a concept deck.

Claude API Gmail API React
Personal 2026
iv Philosophy
Why I build

AI as amplifier.

The shift wasn't learning to use AI. It was realising I could actually fix things that used to just be the cost of the job — research that dies in folders, a handoff that keeps drifting, a job search full of copy-paste. Before, you worked around them. Now I build the fix.

None of these projects started as portfolio pieces. They started because something was slow or broken. The finance tracker is for two people. The Apply Agent is for one job search. The version system came from a wall I actually hit. The most interesting thing about building getting cheap is what you choose to spend it on.

The most interesting question isn't what AI can do. It's what you stop putting up with once building gets cheap.
— Divya Kulshreshtha, 2026
01

Build for yourself first

The Apply Agent is for one job search. The finance tracker is for two people. If I wouldn't actually use it, I don't build it. That's the only filter that's ever worked.

02

The problem shows up when you're building

The version system didn't come from thinking about design tooling. It came from losing work. The research pipeline came from watching synthesis die in real time. You don't see the wall until you're in it.

03

Run it before you write about it

The agentic inbox is a prototype I actually use. The audit bot runs nightly. Writing about something you've only imagined is just fiction with bullet points.

Building something similar?

Always up for comparing notes on what design process looks like now. If you're figuring this out too, let's talk.

Connect on LinkedIn Hear the podcast