Adalat AI (YC F26)

A legal-tech nonprofit dedicated to transforming justice in developing countries through AI solutions that reduce court delays and improve judicial quality

[Adalat AI (YC F26) @Adalat_AI]

6/ Models + FLEURS-RO benchmark + toolkit:

https://t.co/JD3bSsit6G 

https://t.co/84J7grFrUL

Built on MahaDhwani (AI4Bharat) and NVIDIA Nemotron 3.5 ASR

[Adalat AI (YC F26) @Adalat_AI]

1/ If you’ve ever tried to build with open ASR for Indian languages, you’ve probably run into this problem:

The model can recognize the words — but the output still isn’t ready to use.

Punctuation is missing. Numbers are written out as words. Formatting has to be reconstructed. Before you can turn the transcript into a note, form, court order, or other document, you often need to build another layer of post-processing on top.

Today we’re releasing Koyal, a family of open ASR models designed to close that gap in partnership with NVIDIA Nemotron 3.5 ASR.

4 monolingual models (hi, kn, ml, te) at 120M params + 1 multilingual streaming model at 600M.

https://t.co/j3Knr5Cw9h

@nvidia @NVIDIAAI

[Adalat AI (YC F26) @Adalat_AI]

5/ Everything is evaluated with scribe-eval, our diagnostic framework from SCRIBE (Interspeech 2026): per-category error rates for lexical tokens, numbers, and punctuation, with sandhi-tolerant alignment.

https://t.co/mMHoxOwkZE

[Adalat AI (YC F26) @Adalat_AI]

6/ Models + FLEURS-RO benchmark + toolkit:

https://t.co/JD3bSsit6G

https://t.co/84J7grFrUL

Built on MahaDhwani (AI4Bharat) and NVIDIA Nemotron 3.5 ASR

[Adalat AI (YC F26) @Adalat_AI]

5/ Everything is evaluated with scribe-eval, our diagnostic framework from SCRIBE (Interspeech 2026): per-category error rates for lexical tokens, numbers, and punctuation, with sandhi-tolerant alignment.
https://t.co/mMHoxOwkZE

[Adalat AI (YC F26) @Adalat_AI]

6/ Models + FLEURS-RO benchmark + toolkit:

https://t.co/9n4XxUWCRn

https://t.co/84J7grFrUL

Built on MahaDhwani (@ai4bharat) and Nemotron-3.5-ASR-Streaming (@nvidia).

[Adalat AI (YC F26) @Adalat_AI]

5/ Everything is evaluated with scribe-eval, our diagnostic framework from SCRIBE (Interspeech 2026): per-category error rates for lexical tokens, numbers, and punctuation, with sandhi-tolerant alignment.
https://t.co/mMHoxOwkZE

[Adalat AI (YC F26) @Adalat_AI]

A transcription can look accurate overall and still get a crucial word wrong.

Word error rate (WER) tells you how many mistakes an ASR system made. It doesn’t tell you where those mistakes happened.

SCRIBE, our open source evaluation tool, breaks errors down by category: words, punctuation, numbers, and domain-specific terms.

That makes it easier to see what a model struggles with and where to improve it. We walk through an example using Indic conversational medical speech here: https://t.co/E8GvpHisaz

Our team is presenting two oral papers at #Interspeech2026:

Vividh-ASR, on benchmarking robust Indic speech recognition
Tue, 29 Sep · 17:10–17:30 · C3.2
https://t.co/8SUJdy9dMH

SCRIBE, on diagnostic evaluation and rich transcription models for Indic ASR
Wed, 30 Sep · 15:20–15:40
https://t.co/xX32m9DUYX

Proud of Kavya Manohar, Kush Juvekar, Arghya Bhattacharya, Aditya Srinivas Menon, and Kumarmanas Nethil for this work.

At Interspeech? Find Kavya Manohar and Kumarmanas Nethil to chat about low-resource and multilingual speech.

[Adalat AI (YC F26) @Adalat_AI]

On 8 September 2026, an ethical hacking investigation identified vulnerabilities in Adalat AI’s infrastructure. We worked with the researcher (@ni5arga) to investigate and patch them. No court-user data was accessed whatsoever.

Our joint write-up explains what happened, how AI agents were used to connect vulnerabilities, how we mitigated it, and how separating infrastructure access from document decryption helped protect user data.

As AI agents become more capable at finding and connecting vulnerabilities, we need to keep learning from the wider security community. We’ll be launching a vulnerability disclosure program to invite researchers to help us further strengthen our systems.

We’re sharing these lessons to help others in the ecosystem and build dependable technology that public institutions can trust for the long term.

Thanks to Nisarga Adhikary @ni5arga for his investigation and collaboration.

https://t.co/okikT3Rvoe

[Ministry of Electronics & IT @GoI_MeitY]

.@Adalat_AI (YC F26) is now backed by @ycombinator

Serving 11 states & 6,000+ courtrooms with speech recognition, case management & paperless workflows.

Expanding across India and the Global South.

#IndiaAI #JusticeTech

@AshwiniVaishnaw | @JitinPrasada | @SecretaryMEITY

[Utkarsh Saxena @utkarsh__saxena]

(1/2) Honoured to be on @TIME’s 100 Most Influential People in AI 2026. So much of the AI conversation is about GPUs, chips, and frontier models. Public institutions and the Global South rarely get the same attention. Great to see that work get noticed.

https://t.co/nPaiUMKcjJ