Research

Datasets, Benchmarks, Models.

Open research that powers everything we build. Indian language datasets, evaluation benchmarks, bias taxonomies, and domain-specific models — all from Rota Labs.

Visit Rota Labs Collaborate With Us →

Open Research

What We Build in the Open

We believe the best AI infrastructure is built on open, reproducible research. Here's what we publish and maintain.

storage

Datasets

Native Indian language test sets (not translated from English). Code-mixed corpora across Hinglish, Tanglish, Benglish. Indian bias evaluation datasets covering caste, religion, region, and gender. Domain-specific collections for legal, medical, and government text.

leaderboard

Benchmarks

Indic Eval — the open benchmark for Indian language AI across 22 languages and 6 dimensions. Agent trajectory evaluation benchmarks. Fairness and bias leaderboards calibrated for Indian demographics. Reproducible evaluation pipelines anyone can run.

model_training

Models & Tools

Sandbagging detection probes for catching agents that underperform. Trust cascade routing for multi-agent systems. Multi-script entity resolution models. Transliteration-aware matching. Open source evaluators and safety tooling.

Research Projects

Active Research Tracks

translate

Indic Eval

Open evaluation benchmark for Indian language AI. 22 languages, 10K+ native test cases, 6 evaluation dimensions. The standard for measuring how well AI actually works in Indian languages.

→ shield

Kavach

Defense AI research — adversarial-resistant inference, air-gapped deployment, autonomous systems, and mission-critical reliability for defense applications.

→ rocket_launch

Antariksh

Space AI research — satellite operations intelligence, earth observation analytics, and space situational awareness.

→

Foundational Work

Research That Powers the Platform

12 open source packages. 9 research tracks. Every Rotavision product is built on published, reproducible work.

Sandbagging Detection

96% accuracy detecting AI sandbagging using activation probes — catching agents that deliberately underperform on evaluations.

Powers: Guardian, Kavach

Trust Cascade

Stakes-based routing for multi-agent systems — escalating high-consequence decisions to more capable agents or human review.

Powers: Orchestrate

Indian Bias Taxonomy

Bias dimensions unique to India — caste, religion, region, language, and socioeconomic markers that global fairness frameworks miss.

Powers: Vishwas

Multi-Script Entity Resolution

Semantic matching across Indian data sources — Devanagari, Tamil, Telugu, Bengali scripts with transliteration-aware confidence scoring.

Powers: Context Engine

View all publications at Rota Labs →

Contribute

Build With Us

Our research is open because the problems are too big for one team. We're looking for collaborators across datasets, benchmarks, models, and evaluation methodology. Reach us at [email protected].

Dataset Contributors: Help build native test sets in underrepresented Indian languages
Benchmark Collaborators: Add evaluation dimensions, domain-specific benchmarks, or new languages
Model Researchers: Collaborate on bias detection, agent safety, and Indic NLP models
Open Source Contributors: Evaluators, safety tooling, and data pipelines on GitHub

Academic & Research Access

Universities and research institutions can apply for free API access to Indic Eval, datasets, and evaluation tools for non-commercial research. Email us from your .edu or .ac.in address.

[email protected]

Open datasets. Open benchmarks.
Better AI for India.