Ankit Aggarwal

Wrote the code, started the company, now run the product. Fifteen years in, still chasing the same curiosity.

More gets made than anyone can read. I build small tools that cut the firehose to what matters. Taste is the scarce thing now.

the lab

Each one is a bet I built instead of just talking about it.

Made after hours - each a small, working answer to a "what happens when..." question about AI. Built for myself first, then left online. Some are live, some are open source, and the list keeps growing.

01 Cleo Daily open source

the future is hyper-personal - and taste, not volume, is the scarce thing

An open-source engine that reads hundreds of sources a day, filters out 99% of the noise, and prints a hyper-personal magazine you can hold.

Anthropic routines/MCP/headless Chrome/Python/Tailscale

what it taught meSubtraction was the hard part, not the tech. A quiet news day has to make a quiet issue, or you stop trusting it.

Cover of Cleo Daily issue No.1 - an editorial print brief
03 quiethumans live

now that execution is cheap, what you make is the only signal left

A local-first search engine that finds interesting makers worldwide by what they build, not their follower count or marketing noise.

local-first/Gemma/Qdrant/crawler/FastAPI/MCP

what it taught meWhere you spend the model matters more than which model - cheap work at scale, the costly call only on what survives.

The quiethumans search returning makers whose descriptions never use the searched words
04 CopyStack shipped · free

small native software still wins the things the web fumbles

A native macOS app that stacks what you copy and pastes it back one item at a time - built for one, now used by others.

Swift/Accessibility API/notarized

what it taught mePeople don't want a clipboard that remembers everything. They want one they can hand back in order, then forget.

The CopyStack window showing a numbered stack of copied items waiting to paste
05 HowzYourDay live

writing is high-friction; the interface becomes voice and memory

A voice companion you phone at the end of the day - it listens, then remembers you next time.

Cartesia/Mem0/Gemini Flash/Python/Docker

what it taught meJournaling dies because writing is work. Making the front door a phone call - no app, no login - mattered more than any feature.

The HowzYourDay landing page - a warm dusk orb and the line 'How was your day?'
06 Cleo running daily

own the agent, don't rent it

A personal AI agent on a Mac mini that answers from anywhere and costs nothing per call.

Ollama/Gemma/Parakeet/EmbeddingGemma/Tailscale

what it taught meAn agent you can't see is one you stop trusting. The host mattered more than the model.

The Mac mini and router on an orange shelf where Cleo lives
07 Everything Agent running on the robot

everything changes once an agent has a body

An always-on voice agent for a Reachy Mini robot, with a two-tier brain and choreographed movement.

Python/Claude Agent SDK/MCP/Cartesia/Reachy Mini/Raspberry Pi

what it taught meGive an agent a body and the bar jumps: a confident wrong move is worse than a wrong answer. Latency had to come first.

08 LilCleo open source

software that authors its own art

A tiny desktop companion who walks your dock and reacts to what your Mac is actually doing.

Swift/AppKit/Blender/Python/MCP/mocap

what it taught meThe constraint was the unlock. The app only cross-fades PNGs, so a new character is a new folder, not a rewrite.

Brick, the LilCleo desktop companion, rendered in 22 different moods
09 experimentswith.tech issue 01 · jun

the future of this is people in a room, offline

An offline, in-person gathering for builders, curated by what you've actually shipped.

community/static HTML/Flask/Postgres/Fly.io

what it taught meCurate the room, not the content. Who's in it decides whether it's worth showing up.

experimentswith.tech - a coffee cup, the mark of an offline AI gathering
how i work

Start from a friction I actually feel. Build the smallest real version. Keep what it teaches.

Boring, reliable work is code; work that needs taste is the model. I spend the expensive effort where it counts and nowhere else - and I care whether a thing is trustworthy, not just whether it demos. Same instinct on the platform and in the lab.