Verdandi
An open-source document intelligence platform that supports custom field schemas, replaceable large models, and a complete human review workflow for extracting structured data from documents such as invoices and receipts.
Timeline (last 90 days)
Verdandi is an open-source document intelligence platform created by a former EY risk advisory analyst, aiming to replace opaque, usage-billed closed document processing tools.
The Verdandi project began construction in June 2026.
Verdandi was benchmarked against Microsoft Azure Content Understanding on 100 documents, with the test set including 35 FATURA2 clean invoices, 25 CORD v2 real Indonesian receipt photos, 10 Kaggle scanned invoices, and 30 synthetic documents.
In the benchmark, Verdandi's field macro accuracy was 97.0% versus ACU's 96.2%; Verdandi's line-item F1 was 0.958 versus ACU's 0.947.
In the benchmark, Verdandi's total cost was $0.2619 versus ACU's $2.5522, a cost ratio of about 9.7x.
Verdandi uses the GPT-6-luna model, and ACU uses the GPT-5.5 and text-embedding-3 models.
Verdandi's core design philosophy is that 'the VLM is not judge, jury, and executioner,' and the system does not give the vision model full decision-making power.
Verdandi's OCR layer generates structured text information as well as data coordinates, rather than simply dumping large blocks of text.
The Text Oracle provides independent evidence by checking whether the value claimed by the vision model actually appears in the text near the location it claims; if it cannot be found, it raises a flag rather than inventing a substitute value.
Verdandi evaluates confidence at the field level, allowing uncertainty to be attached to specific pieces of information.
1 more timeline events
All Facts (11)
Verdandi's template system allows users to describe documents and add unique vocabulary, and supports an 'Allow-no-match' feature that lets the VLM flag and abstain when uncertain.
50%UnverifiedVerdandi is an open-source document intelligence platform created by a former EY risk advisory analyst, aiming to replace opaque, usage-billed closed document processing tools.
50%UnverifiedVerdandi was benchmarked against Microsoft Azure Content Understanding on 100 documents, with the test set including 35 FATURA2 clean invoices, 25 CORD v2 real Indonesian receipt photos, 10 Kaggle scanned invoices, and 30 synthetic documents.
50%UnverifiedIn the benchmark, Verdandi's field macro accuracy was 97.0% versus ACU's 96.2%; Verdandi's line-item F1 was 0.958 versus ACU's 0.947.
50%UnverifiedIn the benchmark, Verdandi's total cost was $0.2619 versus ACU's $2.5522, a cost ratio of about 9.7x.
50%UnverifiedVerdandi uses the GPT-6-luna model, and ACU uses the GPT-5.5 and text-embedding-3 models.
50%UnverifiedVerdandi's core design philosophy is that 'the VLM is not judge, jury, and executioner,' and the system does not give the vision model full decision-making power.
50%UnverifiedVerdandi's OCR layer generates structured text information as well as data coordinates, rather than simply dumping large blocks of text.
50%UnverifiedThe Text Oracle provides independent evidence by checking whether the value claimed by the vision model actually appears in the text near the location it claims; if it cannot be found, it raises a flag rather than inventing a substitute value.
50%UnverifiedVerdandi evaluates confidence at the field level, allowing uncertainty to be attached to specific pieces of information.
50%UnverifiedThe Verdandi project began construction in June 2026.
50%