Ecosystem

Solutions

Research

News

Blog

Company

English

Go to Redrob AI

All solutions

Developers and technical teams

Inference on your own machine, and the models underneath it.

On-device inference, and the models underneath it.

Redrob Office answers on the device, and Redrob Labs builds the four model families all of it runs on, including the image model that placed third worldwide among open-source models.

Explore Redrob Office

office.redrob.ai

vendor-agreement-fy26.pdf

The agent, on this file

Read on this device

Clause 4.1

This agreement runs for twenty-four months from the commencement date and renews only in writing.

4.1

Term

Clause 9.2

Either party may terminate on sixty days written notice, and on thirty days for unremedied breach.

9.2

Governs

Clause 12.4

Confidentiality survives termination for a further three years.

12.4

Survives

Notice is 60 days, or 30 if a breach is not remedied. Clause 9.2 is the one that governs, not 4.1. Read on this device, so nothing had to be trusted with it.

Inference on the machine

Documents, code and workflow are answered locally. Nothing leaves the laptop to get an answer, which keeps the data local under the Digital Personal Data Protection Act.

Models we build, not rent

Four model families in house. The image model placed third worldwide among open-source models, with its weights published openly.

The router reads difficulty first

A router grades each request before answering it, so a simple question comes back immediately instead of queueing behind a model built for a harder one.

How the work goes

On-device inference, and the models underneath it.

Read the repository where it already is

Client code under a master services agreement cannot be pasted into a chat box. Redrob Office runs inference locally, so the review happens without the source leaving the machine.

Build the consent path once

Under the DPDP Act a notice has to be standalone, name the specific purpose, and make withdrawal as easy as giving. That is a schema and a UI decision long before it is a policy document, and it is cheaper to get right before the obligations bind.

Write the breach runbook now

The Act gives you no delay at all in telling the affected person, and seventy two hours to file with the Data Protection Board. Seventy two hours is not enough time to decide who owns the pager, and a security failure carries a penalty that runs to two hundred and fifty crore rupees.

Read what Labs actually published

The routing work, the layer re-merging results and the retrieval benchmarks are all written up in full, and the weights are Apache 2.0.

What you use

Redrob Office

Documents, code and workflow, with inference on the device. Nothing leaves the laptop to get an answer.

Redrob Labs

The model stack all of it runs on. Four model families, built in house, including the third-ranked open-source image model.

Redrob Chat

Ask in the language you think in, including two languages in one sentence. The agent does not make you pick.

Redrob AI

The agent itself. Twelve Indian languages, code-mixed input, and one place to ask for any of it.

Questions people actually ask

What does the DPDP Act change in my code?

Four things, mostly. Consent has to be recorded per purpose and be withdrawable, so consent becomes a table rather than a boolean. Retention needs a schedule and actual deletion. Data principal rights need an endpoint. And breach detection needs to be fast enough to file within seventy two hours. The Act is in force in stages. That is time you have, and nobody is going to grant you more of it.

Can I build on the models, or only use the products?

Build on them. The image model that placed third worldwide among open-source models is published with its weights under Apache 2.0, which permits commercial use. Redrob Labs publishes the method behind the routing and retrieval work as well, so the results are reproducible rather than only cited.