01 / Evaluation
Models as a system
We continuously monitor and evaluate candidate models and flow configurations, then route every query to the combination that performs best on that sub‑task.
Precision
document
understanding.
Agentic extraction trained on your process, running on controlled, Swiss‑hosted infrastructure. 99.9% accuracy on the workflows that matter.
Accuracy
99.9%
On workflows trained with your own examples.
Hosting
Swiss
Controlled, tenant‑isolated infrastructure.
Models
Routed
State‑of‑the‑art models, selected per task.
Training
~10 ex.
Lift production accuracy from a handful of samples.
— Approach
A small number of ideas, executed precisely.
02 / Understanding
Document‑native reading
Resilience to source‑format change: the system contextualises ambiguous information and holds consistency when documents evolve over time.
2024 Q3 → layout v3
2025 Q1 → layout v4
2026 Q1 → layout v5 handled.
2025 Q1 → layout v4
2026 Q1 → layout v5 handled.
03 / Training
Routines, not prompts
Learn a client‑specific task from a handful of annotated examples. A Swiss‑electricity‑invoice routine rose from 98.0% to 99.9% with 10 examples.
Baseline
98.0%
+10 examples
99.9%
— How it works
From file to typed JSON, in six stages.
01
Secure ingest
Your file enters a private, Swiss‑hosted pipeline: hashed, tenant‑isolated, and cached.
— Ingest
02
Read the page
Several vision methods pull every line, table and layout cue into structured data.
— Read
03
Recognise
A classifier matches the document against your library, picking the right schema.
— Classify
04
Extract
A routine trained on your process reads for the fields you care about.
— Routine
05
Verify
Validation rules and field weights catch edge cases, only the failing stage re‑runs.
— Validate
06
Deliver
Typed JSON arrives at your endpoint via polling or webhooks.
— Return
Companion product · beyond extraction
Divico.
Controlled agentic execution, trained on your process.
When document understanding is step one of a longer workflow, Divico picks up where Vindonissa leaves off. Teach it a plan (in text, or by screencasting the task once) and it executes on demand on Swiss‑hosted infrastructure. Monitored, auditable, and pushed to 99% accuracy by the same routine training that powers Vindonissa.