SavianoSwitzerland

AI applications

We use artificial intelligence where it produces a measurable gain, and we are willing to say where it does not.

Artificial intelligence

Production and translation at scale

Translating a catalogue of several thousand records into five languages is not a literary quality problem, it is a pipeline problem: detecting what must be translated, preserving markup, keeping units and references intact, and automatically verifying that nothing from the source language leaked into the target.

We build those pipelines with automated checks at every stage. The check that matters is not the model score, it is the audit of the final rendering in a browser, language by language.

  • Catalogue translation preserving markup and units
  • Automatic detection of source-language leaks
  • Editorial generation under verifiable constraints
  • Quality measured on the rendering, not on raw output

Document extraction and structuring

A large share of value sits in documents nobody reads: supplier invoices, spreadsheet price lists, purchase orders, bank statements, catalogues in portable formats. Reliable extraction turns a folder of files into a basis for decisions.

We treat those flows with a traceability requirement. Every extracted value keeps a link back to the document and page it came from, which makes an error detectable and correctable.

  • Extraction of prices and terms from supplier documents
  • Bank reconciliation and accounting records
  • Normalisation of heterogeneous catalogues
  • Every value traceable back to its source

Tool-equipped agents, under budget and under log

A useful agent is a bounded agent. We design agents with explicit tools, a maximum budget per task, a log of their actions, and a hard line between what they may propose and what they may execute.

Decisions with financial or legal consequences stay with a person. That is not a token precaution: it is the condition for the system to remain auditable.

  • Explicit tools with permissions checked at execution
  • Budget per task and automatic stop when exceeded
  • Full, replayable log of actions
  • Strict separation between proposing and executing

What we refuse to do

We do not produce content designed to mislead a reader about its origin or its nature. We do not manufacture reviews, testimonials or references. We do not let a model decide an irreversible action on a production system by itself.

These limits are not abstract scruples. They protect the value of what we build: an asset resting on deceptive content loses its worth at the first change of algorithm or regulation.

Division 02

What we deliver

Catalogue translation pipeline

Full translation, automated checks and an audit of the rendering language by language.

Document extraction

Invoices, price lists, spreadsheets and portable documents turned into usable, traceable data.

Internal business assistant

A tool-equipped agent on your own data, with a log, a budget and an explicit scope of action.

Classification and enrichment

Catalogue categorisation, attribute extraction, duplicate detection.

Visual generation

Consistent presentation imagery, with colour control and piece-by-piece verification.

Assessment and framing

An honest measure of the expected gain before development is committed, including the negative conclusion.

Technologies

Language modelsTypeScriptNode.jsPostgreSQLJob queuesImage processing

Let us talk about what you want to build.

Describe your situation in a few lines. If it falls outside what we do well, we will say so immediately.

Get in touch