EkStep Foundation · Bhashini

Anuvaad

The real problem was making sure legal meaning survived translation.

Engineers had already built a raw prototype to prove the underlying translation technology worked. What came to me was a single sentence: we can now translate any document between Indian languages. Turning that into a tool judges could trust meant treating legal translation as a different problem than translation.

The problem

English judgments, a country that reads in twenty-two other languages.

India's Constitution recognises 22 official languages under the Eighth Schedule. Supreme Court and High Court judgments, by constitutional design, are delivered in English alone. For most people affected by a judgment, the reasoning is written in a language they don't read.

What came to me wasn't a requirement document, it was one sentence: we can now translate any document from any Indian language to any other. That sentence hid the real problem, legal translation isn't word-for-word, and technically correct isn't the same as legally correct.

Approach

Three decisions, starting from a prototype and a sentence.

There was no formal spec to work from, and the underlying technology already existed. The product thinking had to originate the scope from a single sentence and a rough prototype.

Decision 01 · Load-bearing

Built the product around a review layer, not the raw translation.

The product was designed as a translation-plus-verification workflow, the engine drafts, subject-matter experts, retired judges among them, review and correct, and those corrections feed back into the system. The tool's job was never to be the final authority.

Decision 02

Designed for legal-domain accuracy, not general-purpose translation.

A generic translator optimises for fluency. A judiciary tool has to optimise for precision, the same word can carry different legal weight depending on context, so the interface was built to surface that specificity rather than hide it.

Decision 03

Built the interaction model working directly against the prototype.

With no implementation requirement to work from, the screens and flows had to be originated from what the technology could already do, plus a single narrative statement of intent.

Outcome

Deployed for two national judiciaries.

Nov 2019

Deployed as SUVAS, launched by then Chief Justice S.A. Bobde.

Feb 2021

The Supreme Court of Bangladesh adopted the same engine as Amar Vasha.

16 languages

SUVAS now supports 16 Indian regional languages, up from nine at launch.

22M+ documents

Digitised and translated across the platform since launch.

Reflection

Handing raw machine-translated output straight to a litigant would have been faster to ship and wrong for the domain. Building the review layer in from the start is why the tool could be trusted with actual Supreme Court judgments, and years of continued use since is the proof it held up. The next problem is coverage, sixteen languages is real progress, still a fraction of what's spoken across the country.