What we build11 of 12Private AI12 things we build
Your files go into public AI tools.
Private AI runs on your own servers and laptops, and your IT decides what it may touch.
Tested on twenty of your own past jobs before it goes live, and kept working as AI changes.
Eashan Gupta, founder. You meet him on the first call.
All 12
Everything we build for private AI.
Each one opens with what we build, three examples and what we don't promise.
A sample company brain: a pricing question answered from your own price sheets and past quotes, with both sources.
-
Your own AI assistant, on your own servers
A private assistant for drafting and reading company files, on servers you control, with every use logged.
What we build
- An assistant your staff open in the browser, on the phone or in the team chat, that reads files and drafts mail.
- Open models running on your own server or private cloud, sized to your team.
- Sign in with your company accounts, with each team seeing only its own files.
- A log of every question and answer, kept for as long as your policy says.
- A switch to a stronger outside model for the jobs you mark as safe to send.
For example, in your trade
- Law firms. A CA firm's staff ask questions of client ledgers and tax notices, and nothing leaves the firm's own server.
- Diagnostic labs. A hospital's doctors summarise discharge notes with an assistant that runs inside the hospital network.
- Agencies. An agency drafts pitches from unreleased client briefs without a word going to a public AI service.
What we don't promise. Open models that run privately trail the best paid ones, and the gap grows the smaller your server, so we don't claim they match and we show you the difference on your own work. On WhatsApp it answers from your own records, as a service of your business. A general assistant runs in your own app, the browser, Slack or Teams, because WhatsApp's business rules have barred those since 15 January 2026.
-
Every job on the right model, with a backup
Each job runs on the model that suits it, and moves to a backup when one fails.
What we build
- Each job sent to the AI model that does it best for the money.
- Automatic fallback to a second provider or your own server when the first one fails.
- Your prompts and tests kept apart from any one provider, so a switch never means starting again.
- Status checks every minute, with an alert when a provider slows down or fails.
For example, in your trade
- Hospitality. A hotel group's booking assistant keeps answering guests when one AI provider goes down, because requests move to the next one.
- Store chains. An online store's product search keeps working through a provider outage on the busiest sale day.
What we don't promise. A backup model may answer a little differently, so we test every backup on your cases and don't promise identical answers.
-
AI that runs where your data has to stay
Models set up in your building, an Indian data centre or your own cloud account, so data stays put.
What we build
- A plan for where each kind of data may go, drawn from your contracts and the rules that apply to you.
- AI servers in your office or data centre, sized for your users, with spares and support.
- Or models in an Indian cloud region or your own cloud account, with keys you hold.
- One checkpoint every AI request passes through, with a map of where each was answered.
- Updates, monitoring and a monthly health report, run by us.
For example, in your trade
- Finance. A lender reads loan files and bank statements with AI that runs in an Indian data centre, under keys it holds.
- Manufacturing. A plant keeps its drawings and process data on an AI server in its own building, with no line to the internet.
- IT firms. An IT services firm sets up a private AI box in each customer's office and runs it under its own name.
What we don't promise. A server in your office is only as safe and reliable as the room it sits in, so we plan power, backups and access with you and don't promise cloud level uptime from one box. The open models closest to the best need a large server; a box for one office runs smaller models, and we show you what that costs in quality before you buy.
-
A model trained on your own work
A model taught on your past work, so it writes, sorts and answers the way your best people do.
What we build
- Your past work cleaned into training examples: good drafts, right answers, correct labels.
- A model tuned on those examples, run on your servers, that you own outright.
- A test set from your real cases, so you see the gain before you switch.
- Retraining every quarter as new examples come in.
For example, in your trade
- Finance. A broker's model sorts incoming claims mail into 30 types the way its senior staff would, and routes each to the right desk.
- Manufacturing. A plant's model writes inspection notes in its own format and terms, trained on five years of approved reports.
- Recruitment. A staffing firm's model writes candidate summaries in the style its clients have approved for years.
What we don't promise. Training teaches a model your style and judgement, not new facts, so answers about facts still come from your live records, and we only train where a test shows it helps. We train only on data your customers were told could be used this way, with names and numbers masked first.
-
Small, fast models for the jobs you run all day
A small model trained for one job, like sorting mail or reading forms, at a fraction of the cost.
What we build
- A study of your AI jobs to find the repetitive ones a small model can carry.
- A small model taught from the big model's best answers on your own cases.
- A side by side test on your work before anything switches over.
- The small model on a modest server, a laptop or a phone, with the big one kept for hard cases.
For example, in your trade
- Moving. A freight firm's small model reads every rate request and fills the quote sheet, and only odd ones go to the larger model.
- Software companies. A software company tags thousands of support tickets a day with a small model on one server instead of a paid AI service.
What we don't promise. A small model is only good at the job it was trained for, so it hands anything unusual to a larger model or a person.
-
AI on phones and laptops, even with no signal
AI that reads, listens and answers on the device itself, so it works offline and nothing leaves the phone.
What we build
- Small models built into your app for reading forms, taking voice notes and answering from a manual.
- Speech to text that runs on the phone or laptop, in the languages your staff speak.
- Work saved on the device and synced when the signal comes back.
- The same features on a laptop for staff who must keep files off the cloud.
For example, in your trade
- Telecom. A telecom contractor's engineers log a tower fault by voice in a basement, and it syncs when they're back in the open.
- Agriculture. An agri company's field officers photograph a leaf and get advice in Marathi with no internet on the farm.
- Diagnostic labs. A home care nurse dictates notes on a phone that never sends the audio anywhere.
What we don't promise. Phone models are small, so they do narrow jobs well and pass anything hard to a server when there's signal, and older phones may not run them at all. Advice comes only from your experts' approved guide, and an unclear photo goes to an expert before anyone acts on it.
-
Your AI bill, watched and cut
Every rupee of AI spend tracked by team and job, with caps, alerts and cheaper routes where quality holds.
What we build
- One checkpoint every AI request passes through, with spend by team, app and customer.
- Monthly caps and alerts per team, with a hard stop you control.
- Repeated questions answered from memory, and bulk jobs run overnight at lower rates.
- Routine work moved to smaller or private models, after a test shows quality holds.
- A monthly report: what was spent, what it bought and what we cut.
For example, in your trade
- Software companies. A software company sees the AI cost of each customer account and prices its AI plan so it never loses money on one.
- Larger firms. A shared services centre charges AI use back to each department every month, with caps set by the CFO.
What we don't promise. We don't promise a saving before measuring your use, and no job moves to a cheaper model until a test on your own work shows the quality holds.
-
Always the best model for each job, tested first
New models are tested on your own cases each month, and switched in only when they're better or cheaper.
What we build
- A test set built from your real work, scored the same way every time.
- A monthly run of new models against it, with cost and speed side by side.
- Switches made with a one-step way back to the old model.
- A one page note each month: what we tested, what changed and why.
For example, in your trade
- Law firms. A law firm's contract reader is tested on 200 of the firm's own clauses each month, and moves only when a new model scores higher.
- Software companies. A software company gets a monthly note comparing new models on its own support questions, with the cost of each answer.
What we don't promise. Public leaderboards don't predict how a model does on your work, so we only trust your own test set, and we never switch without a way back.
-
AI that understands how your customers really write
AI that reads and replies in Hinglish and your customers' Indian languages, tested on their own messages first.
What we build
- A test of the leading Indian-language models on your own customers' messages and documents.
- The winner tuned on your words: product names, local terms and how people actually spell them.
- Reading of forms, letters and notices in Indian scripts.
- Everything run on servers in India.
For example, in your trade
- Finance. A lender understands borrowers' WhatsApp replies in Hinglish and Marathi and logs each promise to pay.
- Government. A state department reads citizen complaints in Odia and English and sends each one to the right office.
- Agriculture. An agri input company answers farmers' questions in Telugu and Kannada from its own crop guides.
What we don't promise. Quality differs a lot between Indian languages, so we don't promise the same accuracy in every one and keep a person on the languages the model reads poorly.
-
A picture model that knows your products and your look
An image model taught your products and style, so new pictures stay on brand and true to the item.
What we build
- A small add-on trained on your product shots, packs and past campaigns.
- Pictures made on your own machine, so unreleased products stay private.
- A library of approved styles your team picks from, with no prompt writing.
- A check that compares each picture to the real product before anyone uses it.
For example, in your trade
- Store chains. A beverage brand makes festive campaign pictures with its exact bottle and label, months before the product ships.
- Agencies. An agency trains one add-on per client, so every concept it shows already uses the client's real products.
What we don't promise. Even a trained model can slip on small text and fine detail, so no picture is used until it's checked against the real product.
-
A private AI for the owner and the board
An AI that has read your plans, numbers and board papers, for testing a decision before you make it.
What we build
- A private space only you and the people you name can open.
- Your plans, accounts, board packs and market notes read into it, on servers you control.
- Answers that show the page and the number they came from.
- Scenarios worked through: what a price rise, a new branch or a lost client does to cash.
For example, in your trade
- Manufacturing. A second generation owner tests what a 10% steel price rise does to each product line's margin before calling buyers.
- Brokerages. A developer's board asks how a six month delay on one tower changes cash for the whole group.
What we don't promise. It works through the numbers you give it and shows its sums, but it isn't financial or legal advice and the decision stays yours.
-
Coding AI that never sends your code outside
AI coding assistants and agents for your developers, on your own servers, so client code never leaves.
What we build
- Code completion and chat inside your developers' editors, served from your own server.
- Coding agents that fix issues and write tests in a walled-off copy of your code, with a person approving every change.
- Rules for each client project on which model may see which code.
- A log of every request, ready for your clients' audits.
For example, in your trade
- IT firms. An IT services firm gives 300 developers code completion that runs in its own data centre, as its banking clients require.
- Finance. A bank's in-house team uses AI coding agents on a server inside the bank, and every change goes through review.
What we don't promise. Private coding models trail the best cloud ones on hard tasks, so we don't claim they match and show you on your own code where a person still does better.
Most owners don't know
Where private AI stands now.
A sample job on two models: a small one reads every supplier invoice first and sends the hard ones up to a large one.
A small model can match the giants at one job
Trained on your own examples, it costs a fraction to run.
Private AI is closing the gap
The best open models come close to the paid ones. We measure the gap on your work.
Useful AI with no internet at all
Phones and laptops run AI with no signal and nothing sent anywhere.
It works today. We keep it working tomorrow.
How it works- Work it out, most consultants stop here
- Build it, most software firms stop here
- Run it
- Answer for it
- We stay
Bring us what can never leave your building.
In 45 minutes with Eashan you leave knowing what we'd build first, and why. Nothing to sign.
What stays yours
- Your accountsin your company's name
- Your codeevery line of it
- Personal dataabout your customers and staff, on your own machine unless you approve otherwise in writing
- Plain files and a how-to guideif you stop
Eashan Gupta, founder. You meet him on the first call.