Document AI

Industrial OCR Engine Audit & Roadmap for Tessi

Advised the Technical Direction of Tessi, a European leader in document processing, on the evolution of their industrial OCR engine: a benchmark against state-of-the-art models, an end-to-end architecture audit and a technical roadmap for the next-generation platform.

SOTA benchmark
End-to-end audit
Next-gen roadmap

The problem

Tessi is a European leader in document processing. Its industrial OCR engine sits at the center of high-volume workflows where documents arrive in every condition: scans, photos, forms, mixed print and handwriting, varied layouts and languages. At that scale, small differences in accuracy turn into large differences in manual correction work, and small differences in cost per page add up quickly.

Document AI has also changed fast. Compact document models and vision language models (VLMs) now report strong results on public benchmarks such as OmniDocBench and olmOCR-Bench, and it is no longer obvious which parts of an established OCR stack should be kept, replaced or combined with newer models. Tessi's Technical Direction wanted an independent, evidence-based view before committing to the next generation of the engine.

Our approach

This was an advisory engagement with Tessi's Technical Direction, delivered through Molia. Their team owns and runs the engine; our role was to measure, challenge and recommend. The work had three parts.

Benchmarking. We challenged the existing stack against state-of-the-art models. A useful OCR benchmark goes beyond a leaderboard run: it needs a corpus that reflects real production documents, metrics that match how the output is consumed (character and word error rates, but also field-level accuracy), and a cost and throughput view per page. Public benchmark scores are a starting point, not an answer, because production documents rarely look like benchmark documents.

Architecture audit. We evaluated the end-to-end pipeline: preprocessing, the OCR models themselves and post-processing. In OCR systems, errors often originate upstream of the recognition model, in image cleanup, deskewing, page segmentation or text detection, and are then either corrected or amplified downstream by post-processing. Looking at each stage separately is what makes it possible to tell where accuracy is actually lost.

Optimization. We focused on the two constraints that matter together at industrial volume: accuracy and scalability. A model that is more accurate on a sample but too slow or too expensive to run on every page does not help a high-volume operation. The question is where newer models earn their cost, and where the existing components remain the right choice.

The output was a technical roadmap for the next-generation platform, grounded in the benchmark and audit results. This combines our computer vision practice with the evaluation discipline of our research & development work.

Results

SOTA benchmark: the existing OCR stack was measured against state-of-the-art models, giving the Technical Direction a direct comparison instead of relying on published claims.
End-to-end audit: the full pipeline, from preprocessing through OCR models to post-processing, was evaluated stage by stage.
Next-gen roadmap: we delivered the technical roadmap for the next-generation platform, covering accuracy and scalability for high-volume document processing.

What it takes to act on an OCR roadmap

Moving an industrial OCR engine to a new generation is an engineering program, not a model swap. It typically involves:

A representative evaluation corpus with ground truth, versioned and kept separate from any data used for tuning, so every change can be measured against the same reference.
Field-level metrics tied to business outcomes, since a single wrong digit in an amount or an identifier matters more than an average character error rate suggests.
Cost and latency budgets per page, set before choosing models, so accuracy gains are weighed against what they cost to run at volume.
Confidence scores and routing, so uncertain pages go to a heavier model or human review instead of passing through silently.
Controlled rollout, with shadow runs on live traffic before any component is replaced in production.

We cover the benchmarking and architecture side in more depth in our guide to benchmarking OCR engines against vision language models.

Where this applies

Any organization processing documents at volume faces the same choices: banks, insurers and asset managers handling statements, forms and KYC files in banking & finance, and operators and public bodies digitizing records and correspondence in industrial & public sector settings.

Evaluating your OCR or document AI stack against newer models? Get in touch.