An SLM, running on edge, as a sovereign stack, with guardrails.

Your data stays with you, the compute runs. Both on your mobile.

A billion data centres.

India has around 700 million smartphones, and each one can run its own compute. Jil runs on the phone you already own.

Two strands, one phone.

Your record and your compute are wound together on your device. The data stands still and the model comes to it, with an encrypted backup only you can open.

It takes the errand in.

Forms, OTPs, bills, crashed upload links. Jil takes in the busywork, sorts it, explains it and shows you the arithmetic. It doesn’t choose for you.

Four public rails, and only four.

Identity, money, health and travel orbit your Jil. When a payment is due, Jil hands you to the UPI app you already use.

Agents meet only when you introduce them.

Millions of skies, each answering to one person. Your agent talks to someone else’s only after one of you adds the other, like a contact.

Scroll

Notes on building personal AI for India.

Pick a frame to step into the gallery.

Drag to turn the orbit

Walk the gallery

    Jil says hi!

    How are we doing today?

    Chart No. 1: A day, given backA story in four pictures

    Our promise

    We, the makers of Jil, having resolved to build a personal agent that sees but does not decide, and to secure to every person who uses it:

    1. Consent, given once or each time, and taken back at will, with the final say always theirs;
    2. Ownership of their own record, kept on their own phone, to see, correct, carry and erase;
    3. Privacy, with nothing told beyond what is asked, and no one let in unless they let them in;
    4. Honesty, of source and of doubt;
    5. Loyalty, to one person, with no second master, and no use of their life to train it without their leave;
    6. Access, on the phone they already have, with another way in when a body or a machine fails, and the freedom to leave;
    7. and Accountability, with Jil’s rules, its receipts, every tool it connects to, and its makers open to independent audit;
    8. and promoting among all who use it Trust, earned one kept promise at a time;

    Do hereby declare these the terms Jil is built on, and give them to ourselves, to be held to by anyone.

    Printed and published by

    Every person gets one sky.

    Jil is the sun on your own phone, and the only thing that sees your whole life at once. Four public rails orbit it the same way for everyone. Comets carry one consented question between skies, then fade.

    Identity

    DigiLocker, your IDs, your registrations. Shown when an errand needs them, never handed over wholesale.

    Money

    UPI, bills, fees, savings. Jil does the arithmetic; the final tap happens in your own UPI app.

    Health

    Records, prescriptions, vaccinations. Kept on your phone, readable by you first.

    Travel

    Trains, buses, metro, flights. Tickets and timings, sorted without another app to learn.

    Guardrails

    An agent that can act for you has to be built so it cannot act against you.

    These are the guardrails Jil is built on, why we think they matter, and where they sit in what India is already building.

    Why an agent needs guardrails

    An app waits to be opened. An agent doesn't. It holds your context across your money, your health, your papers and your plans, and it can act on them. That is what makes it useful, and it is also why the old bargain of the internet, tap "I agree" and hope, does not survive the move from apps to agents.

    A personal agent can fail the person it serves in three ways. It can get something wrong: misread a lab report, quote the wrong EMI, pay the wrong merchant. It can let someone else in, because every add-on or connector is a door, and a door anyone can install is a door anyone can walk through. And it can quietly take on a second master: an advertiser, a lender, or a model that learns from your life for someone else's benefit.

    The second failure is not hypothetical. In early 2026, security researchers audited ClawHub, the add-on marketplace for the popular open-source agent OpenClaw, and found 341 malicious skills among 2,857, most of them built to steal passwords and crypto wallets. OpenClaw ran on people's own machines. That didn't help, because the marketplace was open by default, and running locally protects nothing once you have invited the thief in.

    In India the stakes are sharper. Phones are shared across a family. Millions of people are coming online for the first time in their own language, and for them a wrong payment isn't a support ticket, it is the week's groceries. Household money is often collective, not individual. An agent designed around one person's credit history in one person's pocket will misread all of this.

    Trust in UPI was not a promise made by any one app. It was a property of the rail: the money either arrived or came back, the rules were the same for every app, and a neutral operator stood behind them. That is the model we are following. Guardrails are not a brake on adoption. They are the reason anyone would let an agent near their bank account at all.

    India is building the rails for agents now

    For once, the public infrastructure is being designed at the same moment as the products that will run on it. That is rare, and it is the reason to build here.

    1. The RBI publishes its FREE-AI framework for AI in finance: seven principles and 26 recommendations, which already anticipate agents fetching offers and executing transactions.

    2. MeitY's India AI Governance Guidelines set out seven sutras, from "Trust is the foundation" to "Understandable by design". The DPDP Rules are notified on 13 November, starting an 18-month rollout of India's data protection law.

    3. At the India–AI Impact Summit, Mastercard demonstrates agent payments with Axis Bank and RBL Bank.

    4. Pine Labs' P3P goes live, letting an AI agent complete a UPI payment on a single authorisation. Mastercard completes what it calls its first authenticated agentic transaction, in New Delhi.

    5. Agentic AI is one of three pillars at Global Fintech Fest, where the Prime Minister names it a frontier and calls for a Fintech Consumer Protection Index. NPCI's Unified Agent Protocol, reported in detail that week, would let people delegate small UPI payments to an agent inside limits they set, built on UPI Circle and Reserve Pay, with a registry to verify and monitor agents. Who is liable when an agent pays wrongly is still unsettled.

    6. DPDP Consent Managers can register: neutral intermediaries through which a person gives, reviews and withdraws consent across services.

    7. The rest of the DPDP Act becomes enforceable: notice, consent, security safeguards, breach reporting, and the rights to access, correct and erase your data, with penalties of up to ₹250 crore.

    Underneath the rules, the capacity is arriving. The IndiaAI Mission has put more than 40,000 subsidised GPUs, the AIKosh library of public datasets and models, and several sovereign foundation models into the ecosystem. UPI processed 24.51 billion transactions in August 2026 alone. In EY and CII's survey of more than 200 Indian enterprises, 24% already report active agentic deployment, the leading worries are hallucination and data governance, and small models that run on ordinary phones and patchy networks are named as the practical path for India.

    The market is moving faster than the rules. Agents are beginning to pay before anyone has agreed who pays when they get it wrong. That gap is exactly where guardrails belong, and a product that waits for regulation to fill it will spend its first years on habits it later has to unlearn.

    What the rest of the world is doing

    Elsewhere, guardrails for agents are being written in one of two places: inside private payment networks, or in regulations that keep moving. Visa's Trusted Agent Protocol, Mastercard's Agent Pay and Stripe's agentic commerce tools each bring their own agent identity, their own consent record and their own liability contract. The EU AI Act, the most detailed rulebook anywhere, has pushed its obligations for high-risk systems from August 2026 to December 2027. In the United States, DARPA's MATHBAC programme, launched in April 2026, is funding the basic mathematics of how agents communicate with one another, which tells you how early all of this still is.

    India's approach is different: define the protocol once, at the level of the public rail, so that an agent registered once can work with any bank and any merchant. That is the grain we are building with. A personal agent designed here, against Aadhaar, UPI, the Account Aggregator framework and DPDP consent, starts from guardrails most of the world is still negotiating.

    Our guardrails

    Our preamble names the values. This is what each of them means in practice.

    1. Jil sees. You decide.

      Jil sorts, explains and shows you the arithmetic. It doesn't counsel you or choose for you. If the numbers say you can't afford something, it tells you plainly, and you can still go ahead. The friction scales with how reversible an action is: reading a bill needs nothing from you, moving money needs you.

    2. Money moves only on your say.

      When a payment is due, Jil hands you to the UPI app you already use. It pays on its own only where you have given standing permission, inside limits you set, the same shape NPCI's agent protocol is taking. Money leaving your account should stay reversible by a human for as long as possible.

    3. Your record lives with you.

      Your personal data stays on your phone by default, with an encrypted backup that only you can open. No government rail and no one at Compute can query it. You can see where every fact came from, correct it, carry it elsewhere or erase it. These are the rights DPDP gives you, built in rather than applied for.

    4. Four public rails, and only four.

      By default Jil connects to the rails India already runs: identity (Aadhaar and DigiLocker), money (UPI, and the Account Aggregator framework for consent-scoped views of your accounts), health (ABDM and your ABHA ID) and travel (DigiYatra). Identity is the root the other three rest on. Every person has access to all four, and nobody is required to use any of them.

    5. Nothing else gets in without passing four gates.

      Any third-party tool has to be open source, so its code can be read; crowd validated, so it has actually been reviewed and tested, not merely published; ledgered, with a durable entry anyone can check; and authenticity approved, with its maker's identity verified. The gates tighten as the stakes rise, so a recipe plugin is not held to the standard of something that touches your money or your health. Nothing is exempt, including a tool you add yourself.

    6. Agents meet only when you introduce them.

      Your agent can talk to someone else's only after one of you adds the other, the way you add a contact. Being able to see another agent never gives either of you the right to speak to it.

    7. Consent is given once or each time, and taken back at any time.

      You decide whether a permission stands or is asked for every time, and you can withdraw it whenever you like. This is the same logic as the Consent Managers DPDP brings in from November 2026.

    8. One person, no second master.

      Jil works for the person who uses it and answers to no one else. Nothing about your life is used to train it without your separate, explicit permission.

    9. A way in for everyone, and a way out.

      Jil is designed for the phone you already own. A compact on-device model needs about 1GB of memory, well within the 4GB that even sub-₹10,000 Android phones now carry. When a fingerprint or a face doesn't match, there is always a non-biometric way back in. And you can leave whenever you like, with your record.

    10. Anyone can hold us to this.

      What gets audited is receipts, never your records: the registry proves that a request happened, not what it contained. That is the difference between an audit trail and a surveillance record. We will publish a log of our decisions, name a person you can appeal to, and petition for a government-constituted tribunal, like those India already has for company law and telecom, with the standing to audit anything built on the registry, including us.

    How this meets India's seven sutras

    MeitY's guidelines set out principles rather than product rules. Our guardrails are meant to extend them into a working agent, not to duplicate them.

    SutraWhere Jil answers it
    Trust is the foundationTrust as a property of the architecture: on-device records, receipts not records, a public decision log.
    People firstJil sees and you decide. Final say, and the ability to override, always stays with the person.
    Innovation over restraintGates that scale with stakes, so low-risk tools are not buried in process and high-risk ones are never waved through.
    Fairness and equityBuilt for the phone people already own, with a non-biometric way in and collective household finances in view.
    AccountabilityA named human to appeal to, a published log, and a tribunal with standing to audit the makers.
    Understandable by designEvery fact shows its source, and every estimate shows its doubt.
    Safety, resilience and sustainabilityReversibility-scaled friction on money, gated add-ons, and a small model that runs locally at almost no marginal cost.

    What we haven't solved

    Honesty of source and of doubt applies to us too. These are the questions we are still working on, in public.

    • Who pays when an agent pays wrongly. NPCI hasn't settled this, and neither have we. One principle under discussion is that liability should follow control, and that nobody should have to work out whose system failed before getting their money back. We think that is right.

    • When the model is simply wrong. A misread lab value or a wrong EMI figure is a risk of its own, separate from privacy. Showing the source and the doubt reduces it. It doesn't remove it.

    • Whether a registry becomes a closed shop. "If it isn't on the registry, it isn't in the system" protects people, and it also concentrates power. The tribunal is our answer so far, not a finished one.

    • Whether a ledger becomes surveillance. A record of who connected to what is revealing even when it holds no content. We keep receipts to the minimum. We haven't yet shown that the minimum is enough.

    • When voluntary becomes expected. DigiYatra began as optional and is becoming the norm at some airports. We can promise that our architecture never makes anything mandatory. We can't promise the world around it won't.

    These are commitments, not features. If we break one, we expect to be told.

    Read our promise

    Sources

    1. RBI, Framework for Responsible and Ethical Enablement of AI (FREE-AI), August 2025.
    2. MeitY, India AI Governance Guidelines, November 2025.
    3. Digital Personal Data Protection Rules, 2025 (G.S.R. 846(E)), phased commencement: KS&K explainer.
    4. Reuters via Deccan Chronicle, "India preparing framework to allow AI agents make small UPI payments", 1 September 2026: link.
    5. TNGlobal, "India plans AI-agent registry for UPI payments", 12 September 2026: link.
    6. Startup Fortune on Pine Labs P3P, Mastercard and Visa agent protocols: link.
    7. The Hacker News, "Researchers find 341 malicious ClawHub skills", February 2026: link.
    8. EU Digital Omnibus on AI, high-risk obligations deferred to 2 December 2027: Hogan Lovells.
    9. EY and CII, The AIdea of India: Outlook 2026.
    10. Nikhil Pahwa, "When should AI agents need an identity card?", MediaNama.

    About us

    Imagine if your WhatsApp, Instagram, YouTube and email talked to each other, and worked in the background. What would you do?

    Your personal agent, your data, staying on your phone, and always in your possession.

    Imagine the time you would get back when the agent removes the spam, unsubscribes for you and reminds you that you must call your mother daily.

    We are making India’s personal agents where the compute runs and you stay in control. India has around 700 million smartphones, each with the ability to be a data centre by itself, and we believe personal agents that run on that data centre are the next step for AI in India.

    Its founders are one that knows AI inside since 1994 and the other that understands the user of that technology.

    We want to make it for Bharat and listen to the people of the land, because they should dictate who reads their data; we honestly have no interest in it.

    Why wait for someone from the outside to build it for us, when we ourselves are more than capable and self-reliant.

    What we believe

    The people

    Find us

    We believe a life is lived in common, the way a chit fund or a joint family carries one another’s weight. We believe in the witness, not the ruler: something that shows you clearly and leaves the choice with you. We believe truth has many names, and no single one should decide for everyone. We believe the small burdens of living should grow lighter, so there is time for the procession to the river.

    So we’re building it.

    A personal agent just for you.

    LinkedInX

    Letters from Compute

    Stay updated

    Occasional letters from us as we build: what we’re making, what we’re learning, and when you can try it.

    Your email, your number, or both.