LLM & RAG

Your documents become a knowledge base you can query in plain language, with sourced answers.

AI only has value when it's grounded in your business. I turn your contracts, procedures and records into a base you query in plain language, and every answer cites its sources. No more hours of searching, no more made-up answers. On a simple question, the answer comes back right away; on a complex one, the system searches in several passes, cross-checks what it finds and verifies its answer before giving it to you, instead of inventing. And when it becomes a copilot for your team, each person only sees what they're already allowed to see: the copilot respects your permissions, it doesn't bypass them.

Key facts

2s

for a sourced answer from your documents

10x

documents processed per day, no rekeying

0

made-up answers tolerated in production

What I build

01

Internal copilot & document RAG

Contracts, procedures, records: your documents become a copilot you query in plain language, with sourced answers. On easy questions, a direct answer; on hard ones, the system searches in several passes, cross-checks the documents and checks its answer before showing it. Your team stops searching, and each person only gets answers on what they're already allowed to see.

Internal copilot · Sourced answers · Iterative retrieval · Permissions respected

02

Document reading and extraction

Invoices, contracts, purchase orders, forms: your documents are read by models that understand the layout as much as the text. The right data comes out structured, ready to drop into your tool, even on PDFs that are scanned, stamped or badly framed.

Extraction · Vision · Invoices & contracts · Structuring

03

Model selection and integration

The right model for each task (Claude, GPT, open-source), cleanly plugged into your systems.

Claude · OpenAI · Open-source · MCP

The promise

The right answer, sourced and cross-checked, in seconds.

Before / after

Hours spent searching through your documents

A sourced answer in seconds

Answers you can't verify

Every answer cites its source

The stack

Claude

OpenAI

RAG

pgvector

Qdrant

Embeddings

Reranking

Python

Vision

OCR

Agentic RAG

Straight answers

01

How long does it take to set up a RAG system?

A POC on a subset of your documents takes a few weeks. A full production RAG, with corpus cleanup, evals and deployment, takes weeks to a few months depending on the volume and state of your documents.

02

Can a RAG system hallucinate?

Far less than a bare LLM, but yes, without guardrails. That's why every answer cites its sources, questions outside the corpus are declined, and I measure the rate of correct answers with evals before every release.

03

Do my documents stay confidential?

Yes. Depending on your constraints, the RAG runs on an EU cloud or entirely inside your infrastructure with private models. Your documents are never used to train public models.

04

Which documents can be plugged in?

PDFs, Word files, emails, internal wikis, business databases: anything with text can be indexed. The real issue isn't the format, it's the quality of the corpus: I sort, deduplicate and date everything before indexing.

05

How does the AI answer a complex or multi-part question?

It doesn't jump on the first page it finds. The system breaks the question down, searches your documents in several passes, cross-checks the results, then rereads its own answer to make sure it holds before giving it to you. In practice, fewer partial or made-up answers on the cases that really matter, and an answer always tied to its sources so you can check it.

06

Doesn't an internal copilot risk exposing the wrong documents?

That's the real blocker on this kind of project, and it's a design criterion from the start. The copilot respects your existing access rights: on every question, it checks what the person is allowed to see before answering, instead of freezing permissions once and for all. A salesperson doesn't stumble onto HR files, an intern doesn't surface executive contracts.

07

Do you only read text, or scanned documents too?

Both. The old approach (character recognition plus rules) breaks the moment a document changes format or comes in crooked. I use models that read the layout the way a human would: they spot a total, a due date or a clause even when it's never in the same place. In practice, your supplier invoices, contracts and scanned forms come out as clean data, verified and ready to enter your system. With mandatory electronic invoicing on the way, it's a project that pays off fast.

Contact

Ready to go from demo to production?

Reply within 24 hours · first conversation free, no strings attached.