An independent AI product lab

AI that keeps track, so you can make the call.

We build small, useful AI products for your time, your health and your work, and one for your time off. Each takes something scattered, like the people and promises in your week, what you ate and how you felt, or what employers actually ask for, keeps it straight and gets the next step ready. The decision stays with you.

Mostly for people in their own lives. Often for professionals. Sometimes for the businesses they work in.

Products

Four products, shipped.

Time

JoeHukum

Remember what matters. Follow through.

A personal CRM and executive assistant. Joe keeps track of the people, details and commitments across work and life, and helps you prepare the next step. You send, book or pay in the apps you trust.

For
Professionals juggling work and home
On
iOS, Android, WhatsApp · beta
joehukum.app ↗

Health

NibbleIQ

Make sense of food and how you feel.

A food and symptom diary. Log meals, drinks and how you feel in your own words, then explore the patterns in your own records. A pattern is a place to look closer. It is never a diagnosis.

For
Anyone tracking what they eat
On
iOS and Android
nibbleiq.app ↗

Work

RoleGPS

Career GPS for AI jobs in India.

A free map of what AI roles actually require, read from more than 1,700 real job postings across 18 roles. Rate yourself against a role, see your gaps, and get the shortest path to close them, including what to skip.

For
Students and career switchers
On
Web. Free, no ads
rolegps.com ↗

Time off

Favseen

Your favourite Netflix scenes, one click away.

A free scene-by-scene guide to Netflix. The best scenes of every show and film, in story order, so you can rewatch a season in an evening, follow a character, or jump straight to a moment. Not affiliated with Netflix.

For
Anyone who rewatches
On
Web and Chrome extension. Free
favseen.com ↗

How we think

Four rules we build by.

A language model is good at reading messy input and preparing a next step. It is bad at being the place where facts live, and it should not be the one deciding. Every product here is built around that split.

  1. Records first, model second.

    What a product knows lives in records you can read and correct. The model reasons over them. It does not get to invent them.

    Every number on RoleGPS links to the job posting it came from.

  2. You take the last step.

    The product does the groundwork and stops. Sending the message, choosing the gift and deciding what a pattern means are yours.

    Joe prepares the message or the payment. You send or pay in the app you trust.

  3. Claim only what we can show.

    AI products overpromise easily. We describe what a product does and where it stops, and we drop taglines that promise more than software can guarantee.

    NibbleIQ calls a pattern an association, not a cause.

  4. No new place to go.

    Help should show up where you already are, without one more dashboard to check.

    Joe answers on WhatsApp. Favseen opens Netflix at the right second. Taught writes to the files your coding agent already reads.

From the lab

Experiments, and things built for fun.

Smaller bets, for brands and for ourselves. Some will grow into products. Some will be retired.

  • Brand Context Protocol

    An open format for a brand's voice, vocabulary and rules, so AI agents can read and respect them. BrandGuard checks a draft against it inside Zendesk and Canva.

    For brands and support teams brandcontextprotocol.com ↗
  • Vault Breaker

    A daily prompt-injection game. The AI guarding the secret word runs entirely on your iPhone, with no server. It is how we learn what small on-device models can do.

    For the curious Play on iOS ↗

For developers

Free and open-source tools.

Pieces we took out of the products we run, for people building with AI agents. Taught is in free beta and agent-tool-result-reducer is open source today. The rest are being released under the MIT licence, one at a time.

  • Taught

    One memory for all your coding agents. Correct Claude Code on one laptop, and Codex or Cursor on the next one already knows. The memory is a private git repo you own.

    Free beta macOS · closed source Install from GitHub ↗
  • agent-tool-result-reducer

    Shrinks large tool results for LLM agents without silently losing data. It keeps whole items within a size budget and tells the model exactly what was left out.

    Open source Python · MIT View on GitHub ↗
  • mcp-rest-bridge

    Puts an existing authenticated REST API in front of Claude, ChatGPT and other MCP clients as a remote MCP server. Each call goes out with the caller's own key, so the bridge stores no credentials.

    Releasing soon Python · MIT
  • pytest-frozen-allowlist

    Architecture rules for a codebase that already breaks them. Each rule has an exception list that can only shrink: no new offenders, and a fixed file has to come off the list.

    Planned Python · MIT
  • youtube-timestamped-rag

    A reference app for asking questions of YouTube videos, where every answer links to the exact second it came from.

    Planned Python · MIT

Each one lands at github.com/AryabhattaLabs ↗ as it is released. Ideas, feedback and design reviews are open in the community repo ↗.

Ajeet Singh Kushwaha

Who's behind it

A small, founder-led lab.

Aryabhatta Labs is run by Ajeet Singh Kushwaha, who has spent 20+ years building products across HealthTech, EdTech, SaaS and AI, at Seekify, Seekho, Freshworks, HealthKart and TATA 1mg. LinkedIn ↗

The lab works across Canada and India, and stays small on purpose. That keeps the loop short between a real problem, a shipped product and what the people using it say.

Where someone knows a domain better than we do, we build with them. NibbleIQ is built with NutriMantra, the nutrition practice led by Shruti Marjara Kushwaha.

Get in touch

Try a product, or tell us what's eating your time.

Feedback on something we've built, a problem you think AI should handle better, or a domain you know deeply and want to build in.