Guwahati, Assam
AI that runs inside your building.
We build systems that sit on your own machines and inside your own pipeline — an offline tutor and record system for schools, and production automation for publishers. Nothing we generate reaches a reader until a named person has approved that exact version.
What we do
Two practices, one engineering approach
Both start from the same constraint: the work has to run where the data already is, and a person has to sign off on what comes out.
Practice 01
For schools
An AI tutor, a Content Studio and a school record system that run entirely on one machine on the school's own network, in Assamese, Bengali, Bodo, Hindi and English.
- No connection, no per-question subscription
- No pupil data leaves the building
- Built around the SCERT Assam academic calendar
Practice 02
For publishers
AI automation built into your existing production pipeline: composition, XML, accessibility conformance and alt text, running on your infrastructure against your files.
- Scoped from a baseline measured on your own pages
- Editor approval gate on everything generated
- You keep the source, the models and the output
Practice 01 — Schools
The classroom has a computer. It does not have the internet.
Everything runs on one machine on the school's own network. A teacher opens it in a browser; nothing is installed on any other device. Open a card to see what each part actually does.
01
The tutor
Answers a pupil's question from the indexed curriculum, citing the file and page it came from.
- Separate modes for pupils and teachers, so a lesson plan's expected misconceptions stay with the teacher.
- Detects bilingual SCERT textbooks and uses the publisher's own facing-page translation instead of machine translation.
- Two performance tiers on one interface — from the office PC to a dedicated classroom appliance.
Every answer carries its source. Pupils learn to check, not just to ask.
02
Content Studio
Generates lessons, projects, worksheets and question banks from the curriculum, into a review queue.
- Nothing reaches a pupil until a named subject expert has approved that exact version.
- Editing an approved item returns it to draft. The approval does not carry over.
- Exports Moodle XML and HTML, so a school already running something else is not locked in.
The review gate is enforced in code, not in policy. There is no path that assigns an unapproved item.
03
The record system
Daily roll call, the monthly attendance register, class reports and offline-verifiable certificates.
- Roll call defaults every pupil to present; the teacher marks the absentees.
- Class reports name which question a class got wrong, not a score.
- Guardians see a summary through a revocable access code — no account, no contact details stored.
Counted fairly: a pupil is never penalised for a day the register was not taken.
04
The Assam calendar
The SCERT Assam 2026–27 calendar, transcribed by hand and reconciled against every summary box in the booklet.
- Region-aware: Barak Valley's Durga Puja offset is a first-class case, not an exception a clerk has to remember.
- Sundays are not assumed non-school — flood and bandh losses are made up on them, and the register knows.
- January to March 2027 report as provisional until the public-holiday notification is out.
Not OCR'd, by decision. A misread holiday silently corrupts every attendance percentage downstream.
05
Five languages
Assamese, Bengali, Bodo, Hindi and English, with no hard-coded English anywhere in the interface.
- A live coverage dashboard shows exactly how complete every language is — nothing assumed, everything checked.
- Every Bodo word on the platform comes from a Bodo-speaking educator — never from a machine.
- Built to be translated: new languages join the platform without touching a line of code.
A blank a dashboard reports beats a wrong word a teacher cannot detect.
06
What it runs on
The PC a school already owns, or one appliance for the whole school.
- Entry tier: the existing office PC — ideal for lesson preparation and planning.
- Classroom tier: one dedicated appliance serving every classroom over the school's own network.
- Both tiers run the full curriculum, Content Studio and record system.
Starting is simple: one machine, one school, the first day of term.
How it works
From the first question to the year-end report
One platform carries a school through the whole academic year — teaching, creating and recording, end to end, in the language of the classroom.
Index
The school's own textbooks and curriculum become the platform's knowledge. What the state teaches is what the tutor knows.
Teach
Pupils ask in their own language and get answers grounded in their own syllabus — every answer citing the chapter and page it came from.
Create
Teachers generate lessons, worksheets and question banks in minutes. A named educator approves every item before it reaches a child.
Record
The morning roll call flows into the monthly register, the year overview and a guardian's summary — one record, told at every level.
Curriculum
An AI curriculum from Class 1 to Class 12
Five bands, each ending in something a child has made rather than something they have watched.
Curious Minds
What is a machine that learns?
Young Explorers
Patterns, data and fairness
AI Builders
Building a first working model
Applied AI
Projects with real data
Advanced
Generative AI and governance
What a school needs to start
Entry tier
- The PC the school already owns. No additional hardware.
- Made for lesson preparation, planning and the staff room.
- Full curriculum, Content Studio and record system.
Classroom tier
- One appliance for the school. A single machine, installed once.
- Real-time answers — fast enough for live use inside a period.
- Serves every classroom over the school's own network.
Practice 02 — Publishers
Production automation that runs on your infrastructure
Composition, XML, accessibility and alt text are the four places where an educational list loses the most hours and the most margin. We build automation into the pipeline you already have, on machines you already control, with an editor's approval on anything a model produced.
01
InDesign composition automation
Scripted page-making that applies your styles from tagged content, rather than a compositor applying them by hand.
- Paragraph and character styles applied from the source markup, driven by your own style manifest.
- Figures and tables placed at their cited reference; master page selected from the content type.
- Running heads, folios, side notes and marginal notes generated rather than keyed.
- Round trip: laid-out pages export back to clean XML, so the composed file is not a dead end.
Built as ExtendScript / UXP against your templates. Your compositors keep the file open and keep control of it.
02
XML-first page production
One structured source, every output — print PDF, EPUB, NIMAS and web, from the same file.
- A schema fitted to your list rather than a generic DTD with tags nobody uses.
- Transforms from author manuscript to composition-ready XML, and out again to each delivery format.
- Validation that fails the build on a structural error instead of surfacing it at page proof.
Where you already run a composition system, we automate into it. We do not ask you to replace it.
03
Accessibility conformance
WCAG 2.2 and EPUB Accessibility 1.1 checks run over every page of every title, not a sample.
- Reading order, heading hierarchy, table headers, language attributes and colour contrast checked per page.
- A per-title report that names the failing element and the page it is on, not a pass/fail badge.
- Output shaped for the conformance statement your customer's procurement actually asks for.
An automated check cannot certify a title. It can tell a human exactly where to look, which is what makes the human's time affordable.
04
Alt text at title scale
Draft alt text for every figure, written from the figure and the text that cites it, queued for a named editor.
- Drafted against the surrounding paragraph, so a diagram is described for what it is teaching, not just what it depicts.
- Decorative images marked as decorative rather than given filler prose.
- Extended descriptions drafted separately for figures that need them.
- Nothing is written into the file until an editor approves that exact string. The approval is stored with it.
This is the same review gate as the school Content Studio. It is the part of our architecture we reuse most.
05
Conversion and reflow QA
Automated comparison between the print page and the reflowable output, so silent losses get caught.
- Content diff between source and converted file — dropped notes, collapsed lists, orphaned captions.
- Equation, table and special-character integrity checked rather than eyeballed.
- A queue of differences ranked by severity, so proofing time goes where it matters.
Built after a baseline on your own converted titles. We report what the check catches and what it misses.
06
Custom AI tools, in your network
Models running inside your firewall, evaluated on your files, with an audit trail you can hand to a client.
- Open-weight models served on your hardware where the content cannot leave your network.
- Retrieval over your own style guides, house rules and back list.
- Every generated item carries who approved it, when, and against which version.
- Evaluated against your existing manual output before anything goes into production.
The offline school tutor is the same architecture under a different curriculum. It exists because a school had no connection; it applies wherever content cannot leave the building.
How an engagement starts
You send twenty pages, not a brief
We run the current process and the proposed one over the same real pages, so the gain is proven on your own files before a contract is signed — not promised in a slide.
The baseline is written down first
Hours per page, error rate, rework rate. Quoted against a number you already recognise, so the saving is checkable afterwards rather than asserted beforehand.
You own what we build
Scripts, schemas, transforms and prompts are delivered as source into your repository. No runtime dependency on us, and no per-seat licence on your own pipeline.
Nothing autonomous ships
Every generated string — alt text, description, summary — is held until a named editor approves that version. Editing it returns it to draft.
About
Built in Guwahati, for schools and publishers we know
Across Assam and the Northeast, thousands of classrooms have a computer but no reliable connection — and pupils who learn best in their own language. VIDYAI EDUTECH PRIVATE LIMITED was founded in Guwahati to close that gap: bringing the most advanced AI of our time into the classroom, on the school's own terms and in the school's own languages. The company is registered in Assam, and the work is done here — for schools we know, by people who sat in them.
Offline by design
The platform runs with the network cable out — by design, not as a fallback. Connectivity is welcome; it is never required.
Reviewed by a person
No generated material reaches a pupil or a page without a named expert's approval attached to that exact version.
Vernacular first
Assamese, Bengali, Bodo, Hindi and English are first-class languages of the platform — not translations bolted on afterwards.
Children's data stays put
No pupil contact details, no dates of birth, no accounts for parents. Guardian access is a revocable code, never a sequential record id.
Get in touch
See it running on your own files
A demo takes about thirty minutes — and runs with the network switched off, because that is the point. Publishers: every engagement is scoped against your own pages, so the value is measured before it is promised.
Or write directly — hello@vidyaiedutech.in