Our conviction
Generative AI is moving from experimentation to industrialisation. But between “running an impressive POC” and “shipping to production that holds”, the gap is huge: observability, costs, guardrails, evaluation, data governance.
Our role: bridge that gap with you, avoiding the pitfalls we learned by shipping our own products (HOLOS, ComptaIA) into production.
Use cases we operate
Business copilots
- Accounting: assistant that proposes postings, generates supporting documents, prepares tax returns
- Audit: assisted analytical review, working memo generation, complex document synthesis
- Legal: contract analysis, clause extraction, version comparison
Structured extraction (OCR + LLM)
OCR + LLM pipeline to extract structured fields from complex documents: invoices, contracts, reports, handwritten forms. Human-in-the-loop validation for edge cases.
RAG (Retrieval-Augmented Generation)
Internal knowledge bases queryable in natural language: product documentation, case law, standards, procedures. Embedding, vectorisation, semantic search, contextual generation.
Conversational automation
Workflows triggered by messages (Slack, Teams, WhatsApp Business) with tool-calling on your internal systems: create a ticket, schedule a meeting, run a report.
Our method
1. Use case framing
Not all AI use cases deserve AI. We systematically frame expected ROI, operational cost (tokens, calls, latency), risks (hallucinations, sensitive data), and alternatives (rules, classical ML).
2. Rigorous evaluation
Every AI system we deliver comes with an evaluation dataset with quantified metrics (precision, recall, latency, cost/query). No evaluation, no production.
3. Guardrails and observability
Input and output filters, full logging, drift alerts, documented kill-switch. Every call is traced, auditable.
Technical stack
- LLMs: Claude (Anthropic API and SDK), OpenAI (GPT-5), open-source models via Ollama / vLLM if sovereignty required
- Frameworks: LangChain, LlamaIndex, custom orchestrators
- OCR: Mistral OCR, Tesseract, AWS Textract depending on case
- Vector DB: Pinecone, Qdrant, pgvector
- Observability: LangSmith, Langfuse, custom logging