01
On-device software
A signed, low-privilege agent that runs on the employee's own machine and inspects what is about to be sent, before it reaches the network. The record can stay on that machine.
Atharia Private Limited · India
Everything we build leaves something behind.
Atharia is a product company. We build software in artificial intelligence, data engineering and enterprise systems, and we operate it ourselves. Our first product, Caldris, governs how AI gets used on the machines where the work happens.
The pattern is not in the sand. It is in the plate.
Chapter 01 — The name
A coined name, from two older words we liked for the same reason.
We wanted a name about consequence — about what is still standing after the work is done. Most software is temporary. We are not interested in building that kind.
athar أثر
Arabic
The trace. The mark, imprint or effect a thing leaves behind once it has passed. Its plural, āthār, is what a civilisation leaves standing after the people who built it are gone.
atharvan
Sanskrit
The keeper of applied knowledge — the tradition concerned not with ritual but with the practical arts: medicine, craft, the working of everyday things. Knowledge that has to survive contact with reality.
Move across the plate. It does not heal.
Chapter 02 — What we build
Four pieces of engineering. One product.
Everything below is built in-house and ships inside Caldris. None of it is sold separately and none of it is sold by the hour — Atharia is a product company.
02
Policy engine
Rules written in ordinary language and compiled to a versioned file. One rulebook covers every surface, instead of one tool for the browser, another for the cloud and a third for email.
03
Signed event log
Every action written as a structured, signed record carrying the finding, the rule version and the device. Built to be read by an auditor in five years, not just a dashboard today.
04
Local detection
Classification that runs on the machine itself, so text can be judged as sensitive without being sent anywhere to be judged. The same models cover typed input and the AI assistants beside it.
We do not sell hours. Atharia sells software on a subscription. Adding the next customer does not mean adding another person to a payroll — which is the whole reason a small team can serve a large number of them.
Chapter 03 — The product
One product, built in the open.
Atharia builds and operates its products directly. Today there is exactly one, and we would rather say that plainly than pad a page with things that do not exist yet.
In prototype
Caldris
Governing AI where it actually happens.
Staff paste customer records, credentials and internal documents into ChatGPT and tools like it. Security software built for email and files cannot see an AI chat, and it cannot see the AI assistants now running on the same laptop. There is nowhere to set one rule and see what happened.
Caldris is software that runs on the computer itself. You write a few rules in ordinary language; when someone is about to send something private to an AI tool, Caldris warns or blocks it and keeps the record on the device.
- The record can stay on the machine. Nothing has to be routed through a vendor cloud, so there is no third party holding your prompts.
- One rulebook, both kinds of use. The same rules cover what a person pastes and what the AI assistants on that laptop do — rather than three separate tools that each see one surface.
- One person can start. A subscription and an install, not a procurement cycle and an IT project.
It is a working prototype and we are pre-revenue. We say so here, and we will keep saying so until the first paying customers change it.
What comes next is deliberately unannounced. We start a product only where a problem someone already pays to work around meets a technical position we can hold — and we will name the second one when it is real, not before.
Chapter 04 — How it works
We moved the control point.
Most tools watch one surface: the browser tab, or a company cloud, or a single vendor's suite. By the time traffic reaches a gateway the disclosure has already happened — it happened on the laptop. So the control belongs on the laptop.
Path of a single prompt
Composing
Step 01
Inspected locally
A signed, low-privilege agent reads the prompt on the machine that produced it, before any network call exists to intercept — and reads what the AI assistants on that machine are doing too.
Step 02
Judged against your policy
One rulebook, written in plain language and compiled to a versioned file. The same rules cover a person pasting and an assistant acting — not three tools that each see one surface.
Step 03
Recorded as evidence
A signed, structured event carrying the finding, the rule version and the device. It can stay on that device — there is no requirement to ship it to us.
Nothing has to route through us, which removes us as a breach surface entirely. And because the product is software rather than a service, one more customer means one more subscription, not one more person on a payroll — so a small team can serve a large number of them, and the engineering roles it creates are skilled ones.
Chapter 05 — Method
Six rules we hold ourselves to.
Not values on a wall. These are the arguments we have already had, settled, and now refuse to reopen under deadline pressure.
Compass and straightedge — the same two tools
i
Start where the problem is
Not where the technology is convenient. If a database serves better than a model, we ship the database and say so.
ii
The customer's data stays the customer's
We design so there is nothing on our side worth breaching. A constraint accepted before writing code, not a policy published after.
iii
Every action leaves a record
Signed, structured, and readable by someone who was not in the room. A screenshot is not evidence.
iv
Small teams, whole systems
The people who design it operate it. Nobody hands a specification over a wall and stops caring what happens next.
v
Build for the maintainer
Software is read far more often than written. We optimise for whoever inherits it — including when that person is us, two years on.
vi
Name what we have not solved
Gaps go on the roadmap in writing, with a date. Overstating coverage is how trust is lost once and never regained.
Chapter 06 — In short
An engineering company that intends to still be here.
Atharia builds applied software: products where a model, a policy or a pipeline is doing real work rather than decorating a demo. One product, Caldris, is in prototype. It is sold as a subscription rather than as consulting hours, which is what lets a small team serve many customers. We are pre-revenue, and the next step is the first of them.