Revolutionary technology redefines visa processing efficiency, ETTravelWorld

The image is used for representation only.

Travel technology startup Atlys has developed BoltOCR, a proprietary 3-billion-parameter vision language model designed specifically for passport extraction, as the company moves away from large, general-purpose language models toward specialized AI systems designed for visa processing.

The internal model shows field-level accuracy of 99.08%, beating GPT-4o by 4.5 percentage points, while processing documents 3.5 times faster at an average latency of 2.3 seconds. Document-level accuracy is 93.6%, which represents a 14 percentage point improvement over GPT-4o.

According to Atlys, the shift to proprietary technology was driven by limitations of existing solutions. “Even a simple mistake – such as a misplaced letter in the name or reading the date of birth incorrectly – can derail the entire visa process,” the company stated. “For Atlys, passport scanning is not just another benefit, but a critical component to ensuring timely delivery of visas.”

Testing revealed systemic vulnerabilities

Prior to developing BoltOCR, Atlys tested enterprise OCR tools, open source models, and leading AI systems including GPT-4o and Claude Sonnet 4. The evaluation revealed consistent shortcomings across platforms: clunky data extraction, slower processing speeds, high costs, and difficulty handling passports captured in real-world conditions with varying quality. “They’re hallucinating. They’re slower. And they’re expensive.”

Purpose built architecture

BoltOCR is designed as a vision language system that focuses on reading passport fields faithfully as they appear, avoiding common errors such as confusion between places of birth and places of issue. The model includes capabilities to normalize inconsistent date formats, parse long or multipart names, and validate entries against the machine-readable zone (MRZ).

A custom quality layer scans documents before OCR analysis begins, identifying unusable inputs such as blurry MRZs, cropped images, reflections, or obstructions. Only valid, processable images make it into the form, preventing computational resources from being wasted on unreadable documents.

The system also adapts to country-specific regulatory requirements, applying strict quality limits where authorities require the quality of original documents while allowing greater tolerance where minor defects are acceptable if the text remains machine-readable.

Cost and operational benefits

BoltOCR runs 4.3 times less expensive than GPT-4o. While third-party API services charge per document scanned, BoltOCR’s internal pipeline operates at largely fixed costs, allowing margins to expand as processing volumes increase.

In addition to cost savings, full inference suite control provides predictable latency and uptime, eliminating potential external interruptions or delays that could impact visa application processes. “Low service costs allow us to keep visa fees competitive, while consistent uptime ensures applications are processed without delay,” Atlis stated.

Expansion beyond passports

While BoltOCR initially focused on passports, its architecture extends to other travel and identity documents including visas, residence permits, and national ID cards. Each type of document presents unique challenges – visas may contain errors from manually filling out forms or embassy processing, while residence permits and ID cards vary widely in format across jurisdictions.

Atlys integrates BoltOCR with document verification layers that ensure that extracted data is not only readable but also compatible with applicant records, creating a unified pipeline across multiple document types.

Benchmarking and future development

At launch, Atlys benchmarked BoltOCR against GPT-4o and Claude Sonnet, which were among the most powerful models available to the public at the time. The company has not yet tested newer models such as GPT-5, which was not released during initial development, but plans to expand the measurement to include newer models in future releases.

Atlys has no plans to release a technical paper or open source elements for BoltOCR, emphasizing that the technology represents a core competitive advantage. The company continues to measure performance against leading models and incorporate the latest research internally, with improvements flowing directly into production systems.

“LLM programs are great for general-purpose assignments,” Atles noted. “But the next era of AI will be about building models tailored to specific use cases – where accuracy, speed and cost are most important.”

This development reflects a broader trend among travel technology companies that are building specialized AI infrastructure tailored to industry-specific needs rather than relying solely on large, general-purpose language models.

Join a 2M+ community of industry professionals.

Sign up for our newsletter to get the latest insights and analysis in your inbox.

Everything about the ETTravelWorld industry directly on your smartphone!