Post 2: The Brain of Art Revisionist – AI Analyzing Documentation

English:

From Database to Thinking System

In my previous post I told you how we gathered all Valsuani information into a structured system. But a collection of documents, however well organized, isn’t understanding yet. We had data, but we wanted insight.

The question was: can we build an AI system that doesn’t just store information, but also understands it, makes connections, and detects contradictions?

The answer: yes. But not the way you usually use AI.

The Problem with Standard AI

If you ask ChatGPT “Who founded the Valsuani foundry?”, you get an answer based on its training data – which is full of the incorrect information we’re trying to correct. AI systems are brilliant at summarizing what’s already known, but less good at discovering that what’s “known” is actually wrong.

We needed something different: an AI specifically trained on our primary sources, not on the secondary sources that perpetuate the error.

Knowledge Base Architecture

We built what you might call a “domain-specific AI” – a system that:

1. Analyzes Documents
– OCR scan of historical documents (handwritten birth certificates, typed business letters)
– Automatic extraction of names, dates, places, relationships
– Classification of reliability level per source (primary vs secondary vs tertiary)
2. Makes Connections
– “Carlo Valsuani died in 1886” + “Claude was 10 years old at father’s death” = Claude born in 1876
– Cross-referencing between family documents and business archives
– Timeline construction: who could have been where, when
3. Detects Contradictions
– Auction catalog says “Marcello founded foundry in 1900”
– But no French business register mentions “Marcello” for 1900, 1910, 1920…
– AI flags this as “claim without primary source”
4. Weighs Evidence
– Primary source (birth certificate) = high evidential value
– Secondary source (museum catalog) = medium evidential value
– Tertiary source (blog without citations) = low evidential value

What AI Can and Cannot Do

This is where it got interesting. AI proved exceptionally good at tasks that are boring and error-prone for humans:

AI excels at:
– Recognizing patterns across 500+ documents (“this name appears nowhere”)
– Spotting timeline inconsistencies (“person X cannot be in two places at once”)
– Tracing citation chains (“these 50 sources all cite one incorrect 1971 catalog”)
AI struggles with:
– Understanding context of historical ambiguities
– Handwritten documents in poor Italian from 1880
– Cultural nuances (why would someone confuse “Marcel” with “Marcello”?)

The solution: human expertise for context, AI for scale.

The Breakthrough Moment

Somewhere in the second month of development, something special happened. We entered a new document – a business letter from 1908 stating “Claude Valsuani, fils de feu Carlo” (son of the late Carlo).

The AI not only recognized the family relationship, but also noted:
– “Carlo” mentioned here as deceased (“feu” = late)
– Carlo’s death certificate: 1886
– Claude would have been 10 years old at father’s death
– No documents suggest Carlo had “Marcello” as a second name

The AI concluded what people had overlooked: if Carlo was the father, and Carlo died in 1886, and there are no documents mentioning “Marcello”, then “Marcello” is probably a later invention.

This wasn’t a magical AI revelation. This was systematic reasoning based on chronology and absence of evidence. But performing this reasoning across hundreds of documents? That’s where AI shines.

Technical Stack

For those interested, our setup:

– Document Processing: Python + Tesseract OCR for handwritten documents
– Knowledge Base: Structured database with relationships between entities
– AI Analysis: Custom GPT-4 fine-tuning on art historical attribution logic
– Conflict Detection: Rule-based system checking claims against primary sources
– Timeline Engine: Chronological consistency checker

From Analysis to Synthesis

The knowledge base became not just a repository, but an active reasoning system. Ask a question, and the system:

1. Searches relevant documents
2. Weighs evidential strength
3. Detects contradictions
4. Presents conclusion with confidence score

Question: “Who founded the Valsuani foundry?”
Answer: “Claude Valsuani (son of Carlo, 1876-1923). Confidence: 98%. Based on: 23 primary documents. Conflicting claims: ‘Marcello Valsuani’ appears in 47 secondary sources without primary support.”

Next Step: Unleashing Multiple AIs

We now had an AI that could analyze our collected knowledge. But what if we didn’t just use our own documents, but searched the entire internet for conflicting information?

That became the next phase: multiple AI systems working in parallel, each with a different angle, each with different online sources. And what they found was… surprising.

The Brain of Art Revisionist – AI Analyzing Documentation