A Marathi-first content engine for the state’s information department.
How we built a production platform for the Directorate General of Information & Public Relations, Government of Maharashtra — turning meeting recordings, official notes and scanned documents into publication-ready Marathi articles, posters, translations and explainer videos.
- CLIENT
- DGIPR, Government of Maharashtra
- IN PRODUCTION
- 10 modules, daily use
- LANGUAGES
- Marathi · English · Hindi
- STACK
- Next.js · Fastify · Postgres
Marathi is the working language — not a translation target.
DGIPR is the communication arm of the Government of Maharashtra. Its District Information Officers carry a continuous publishing load: departmental meetings must become news articles, schemes must become citizen-facing posters, and every piece must ship in Marathi, English and Hindi across print, web and social — accurately and in the department’s established voice.
Generic AI content tools couldn’t do it. They think in English, transliterate names afresh every time, let model knowledge leak into official facts, and mangle Devanagari conjuncts in generated images. Government communication carries accountability: names, dates, amounts and designations may never be invented.
From a meeting recording to a published post.
Every source, reviewed
Recordings are transcribed to Marathi, scanned PDFs go through OCR, and each source is shown page by page so officers correct names and amounts before anything is written.
House-style Marathi articles
Key points, names and designations are confirmed first. Retrieval over the Mahasamvad archive teaches structure and register — never facts.
Posters that fit every point
Templates are chosen by capacity so no point is dropped. Devanagari is typeset by a real shaping engine and the state emblem is composited at exact coordinates.
Translated, checked, published
English and Hindi with a mandatory name check, precision proof-reading, letterhead PDFs and two-step publishing to the department’s official X and Facebook accounts.
Officer in the loop at every stage — and code, not the model, has the final word on facts.
- SOURCESOfficer inputAudio · PDF · YouTube · notes
- STAGE 1Intake & OCRMarathi transcription
- STAGE 2Officer reviewPoints · names · designations
- STAGE 3GenerationStyle RAG over the archive
- OUTPUTSPublishArticle · poster · video · EN/HI
| # | STAGE | WHAT THE OFFICER REVIEWS |
|---|---|---|
| 1 | Source intake | Which pages of a scanned document are relevant |
| 2 | After extraction | Every transcript and page of text, correctable in place |
| 3 | Before writing | Key points, names, designations, editorial direction |
| 4 | Before translating | The exact spelling every name will receive |
| 5 | After production | Article, poster, caption and video — each with its own revision loop |
| GATE | Before publishing | A two-step confirmation on any outward-facing action |