bharath.ai

Second generation of IndicTrans models covering all 22 scheduled Indian languages with state-of-the-art translation quality. Trained on curated parallel corpora including Bharat Parallel Corpus Collection (BPCC). Achieves significant improvements over Google Translate and prior versions on Flores-200 benchmark across all 22 languages.

Transformer-based encoder-decoder architecture trained on large-scale curated parallel data. Uses language-specific tokenization via IndicBERT. Multi-stage training with progressive data scaling and back-translation augmentation across 462 language pairs.

State-of-the-art BLEU scores across all 22 Indian languages on Flores-200. Outperforms Google Translate and prior IndicTrans2 on low-resource language pairs (Bodo, Dogri, Maithili). Open-sourced model weights and training data for research community.

Enables real multilingual AI services across India. Critical infrastructure for government digital services, healthcare, education, and commerce in regional languages. Foundation for India's linguistic AI sovereignty — no dependency on Western models for Indic NLP.

Open platform enabling LLM-powered autonomous software engineering agents. Supports multiple agent architectures including CodeAct for interactive code execution. Benchmarked extensively on SWE-bench and real-world software engineering tasks with contributions from Indian researchers.

Modular architecture supporting multiple agent implementations (CodeAct, Browsing, Delegator). Event-driven runtime with sandboxed Docker execution. Evaluation on SWE-bench Lite and custom software engineering benchmarks with diverse LLM backends.

Achieves competitive performance on SWE-bench with open-source models. CodeAct agent resolves 36.8% of SWE-bench Lite issues using GPT-4o. Framework enables rapid prototyping of new agent architectures. Community adoption: 67K+ GitHub stars.

Provides Indian developers and institutions an open platform for building AI coding assistants. Reduces dependency on proprietary tools. IIT Bombay contributions demonstrate India's growing role in frontier AI engineering research.

Release of the largest open Indian multilingual speech corpus: 10,496 hours across 22 Indian languages from 10,496 speakers. Designed specifically for training text-to-speech (TTS) systems that work across India's linguistic diversity. Built on IndicVoices dataset with refined annotation pipeline.

Large-scale data collection from diverse demographics across India. Automated and manual quality annotation pipeline. Speaker-level metadata including age, gender, region, dialect. Evaluated by training multi-speaker TTS models (VITS2-based) and measuring MOS scores.

10,496 hours of high-quality speech data across all 22 scheduled languages. TTS models trained on this data achieve MOS scores of 3.8-4.2 across languages, approaching human parity for Hindi and Tamil. Largest open Indian speech dataset by 4x margin.

Enables development of natural-sounding AI voice assistants in every Indian language. Critical for accessibility (voice-based interfaces for low-literacy users), government services via IVR, and India's smart city infrastructure. Democratizes speech AI beyond English.

Novel approach to 6DoF pose estimation for previously unseen objects using foundation model features. Eliminates need for object-specific training by leveraging general visual features from DINOv2 for pose regression. Evaluated on standard benchmarks with competitive results against methods requiring object CAD models.

Two-stage approach: coarse pose estimation using PnP with foundation feature correspondences, followed by pose refinement via feature-metric alignment. Uses DINOv2 features for robust matching across viewpoints. No object-specific training required.

Competitive with fully-supervised methods on LINEMOD and YCB-Video benchmarks despite requiring zero object-specific training. Outperforms prior unseen object methods by 8-15% on novel object categories. Real-time inference at 12 FPS on single GPU.

Advances India's robotics AI capabilities — critical for manufacturing automation (Make in India), warehouse robotics, and agricultural harvesting robots. IIIT Hyderabad's Robotics Research Center among top 5 in Asia for pose estimation research.

Contrastive learning-based biomedical foundation model pre-trained on PMC-15M: 15 million biomedical image-text pairs from PubMed Central. Achieves state-of-the-art on medical image classification, retrieval, and VQA tasks. Enables zero-shot medical image understanding.

CLIP-style contrastive pre-training on curated PMC-15M dataset (15M image-text pairs from PubMed Central). Domain-specific tokenizer for medical terminology. Evaluated on 22 biomedical benchmarks spanning pathology, radiology, ophthalmology, and dermatology.

State-of-the-art on 22/22 biomedical benchmarks evaluated. Zero-shot classification accuracy of 89.3% on chest X-rays (CheXpert), 85.1% on pathology (PCam). Outperforms general-purpose CLIP by 15-40% on medical tasks.

Directly applicable to India's AI diagnostics push — enables startups like Qure.ai and Niramai to build better diagnostic tools. With 800M Indians lacking specialist access, AI-assisted diagnosis is infrastructure, not luxury. IIT Kanpur's contribution strengthens India's medical AI ecosystem.

Novel QA system designed for India's National Assessment and Accreditation Council (NAAC) framework. Addresses the unique challenge of extracting and answering questions from Indian educational institution documents spanning multiple formats and languages.

RAG-based (Retrieval-Augmented Generation) architecture fine-tuned on custom dataset of 2,500+ NAAC self-study reports. Uses domain-specific chunking strategies for institutional documents. Evaluated against human expert annotators on 500 real NAAC queries.

Achieves 78.4% accuracy on complex NAAC queries (vs 84% human expert baseline). Reduces institutional self-study preparation time by estimated 60%. Successfully handles bilingual (Hindi-English) institutional documents.

Directly impacts India's higher education quality assurance. 45,000+ colleges undergo NAAC accreditation — AI assistance can dramatically reduce administrative burden and improve quality of self-assessment. First AI system purpose-built for Indian educational governance.

Novel framework for training generalizable robot manipulation policies using semantic data augmentation. Achieves 12x sample efficiency improvement over prior methods. Evaluated on real-world tabletop manipulation tasks with 50+ object categories.

Combines semantic augmentation (generating diverse training scenes via automated visual perturbations) with action chunking transformers (predicting multi-step actions). Trained on 7,500 real-world demonstrations across 12 task families. Evaluated on novel objects not seen during training.

Successfully generalizes to 50+ unseen objects across 12 manipulation tasks. 12x sample efficiency over baselines. Real-world deployment on Franka Panda robot achieving 82% success rate on novel objects. Open-sourced codebase and trained models.

Demonstrates India's growing strength in embodied AI research. IISc Bangalore's Robotics Research Centre collaboration with CMU positions India in the frontier of manufacturing automation AI — critical for Make in India's next phase.

Open-source medical LLMs fine-tuned from Mistral-7B on PubMed and medical textbooks. Achieves competitive performance with GPT-3.5 on medical QA benchmarks while being fully open and deployable locally — critical for Indian hospitals with data sovereignty requirements.

Continued pre-training of Mistral-7B on curated medical corpus (3B tokens from PubMed, medical textbooks, clinical guidelines). Fine-tuned on medical QA datasets (MedQA, PubMedQA, MedMCQA). Evaluated on 10 medical benchmarks spanning clinical reasoning, drug interaction, and diagnosis.

Achieves 72.1% on MedQA (vs GPT-3.5 at 73.6%), 78.4% on PubMedQA. Fully open-source and runnable on consumer GPUs. Model weights and training pipeline released for community use.

Enables Indian healthcare startups and hospitals to deploy medical AI without sending patient data to US cloud providers. Addresses critical data sovereignty concerns in Indian healthcare. IIT Delhi contributors demonstrate India's medical AI research capabilities.

First dedicated Hindi instruction-tuned LLM built on top of OpenHathi base model. Achieves state-of-the-art performance on Hindi NLP benchmarks while maintaining strong cross-lingual transfer to other Indic languages. Trained on 420K Hindi instruction-response pairs.

Instruction tuning of OpenHathi-7B (Hindi-centric base model) on curated dataset of 420K Hindi instruction-response pairs spanning 50+ task categories. Uses DPO (Direct Preference Optimization) for alignment. Evaluated on IndicNLG, IndicNLU, and custom Hindi reasoning benchmarks.

State-of-the-art on 8/10 Hindi NLP benchmarks. Outperforms GPT-3.5 on Hindi comprehension tasks by 12%. Strong zero-shot transfer to Marathi (85% of Hindi performance) and Gujarati (78%). Open-sourced model weights and training data.

Establishes a template for building instruction-tuned models for India's languages. The 420K instruction dataset is itself a major contribution — the largest publicly available Hindi instruction dataset. Directly enables Hindi-first AI products for 528M speakers.