We are introducing ARVA, a multimodal model with a focus on document understanding and optical character recognition (OCR), particularly for Nepali and Devanagari text.
While OCR has seen significant progress for Latin scripts, performance on Devanagari remains inconsistent, especially in real-world conditions. Variability in handwriting, complex character composition, and structural features such as the shirorekha introduce challenges that are not well addressed by existing systems.
ARVA is designed to operate effectively under these conditions.
The challenge
Devanagari OCR is not simply a matter of recognizing individual characters. Characters often connect through a continuous headline, vary significantly in handwritten form, and can be affected by noise, distortion, and low-quality inputs.
In practical settings, documents are rarely clean. They may be scanned at low resolution, captured under uneven lighting, or contain overlapping marks and artifacts. Handwritten inputs introduce further variability.
These conditions require systems that can move beyond isolated character recognition and instead reason over structure, context, and visual patterns.
What ARVA does
ARVA approaches OCR as part of a broader multimodal understanding problem.
Rather than treating text extraction as a standalone task, the model processes visual and textual signals jointly, allowing it to better interpret complex layouts, ambiguous characters, and contextual dependencies.
This enables more reliable performance across:
- Printed Devanagari text under varied formatting and quality conditions
- Handwritten Nepali text with high variability
- Mixed-content documents combining structured and unstructured elements
The system is designed to remain robust in real-world scenarios, where inputs are often incomplete or noisy.
Why this matters
Access to reliable OCR is a foundational capability for many systems, including document processing, identity verification, and digital record management.
In regions where Devanagari is widely used, limitations in OCR systems create friction across both public and private infrastructure. These limitations are often handled through manual processes, reducing efficiency and introducing inconsistency.
Improving OCR performance in these contexts enables more reliable downstream systems and reduces reliance on manual intervention.
System approach
ARVA is built to handle variation and ambiguity as first-class conditions.
Instead of relying solely on fixed pipelines or rule-based corrections, the model evaluates patterns across the entire input, using context to resolve uncertainty and improve recognition accuracy.
This allows the system to better handle edge cases, including partial occlusion, inconsistent handwriting, and structurally complex text.
Looking forward
We see ARVA as part of a broader effort to improve foundational capabilities for language and document understanding in underrepresented scripts.
While progress in OCR has been significant globally, many languages and writing systems still require focused work to reach the same level of reliability.
We plan to continue improving ARVA and exploring how multimodal systems can better handle these challenges over time.