Software and AI engineering · France
Software and AI systems, engineered to run in production.
Cognitum Technologies is a software engineering company. We design, build and evaluate AI systems inside the software our clients run: LLM applications, agents, tool integrations, and the architecture and development process around them. Private models and anonymization when the data requires it.
Capabilities
Architecture, LLM, agents, AI systems, evaluation, MCP
Stack
.NET and Azure, OpenAI Agents SDK, Microsoft Agent Framework, open models
Hosting
Cloud, Azure France, private or on-site when required
Team
Senior engineers, more than twenty years of professional practice
01 — Where we start
Built for teams that need AI to work, not to demo.
A product that needs AI features, a process that needs automating, or data that cannot leave your control: three starting points, one engineering team.
Situation A
Your product needs AI capabilities
You have a software product or an internal platform and need LLM features, agents or tool integrations built by engineers who ship production code, not prototypes.
Situation B
A process is slowing the business down
Quote requests, incoming leads, documents nobody has time to read: repetitive work that AI can take over, with a person keeping the decision.
Situation C
Your data cannot leave your control
Confidential pricing, personal data, regulated sectors: the system must run where the data is allowed to be, with anonymization before anything else.
02 — Capabilities
Six engineering capabilities, one team.
Architecture and development process consulting
Independent reviews of your cloud, web and mobile architecture and of the way your team delivers: pipelines, testing, review, AI-assisted development.
LLM engineering
Retrieval (RAG), embeddings, structured extraction, context engineering and model selection, including open and private models.
Agent engineering
Agents that plan and act through tools, built with the OpenAI Agents SDK and the Microsoft Agent Framework, with approvals where the risk requires.
Building AI systems
The whole system around the model: pipelines, anonymization, orchestration, hosting, monitoring and run.
AI evaluation
Evaluation datasets, metrics agreed with the business, regression gates on prompt and model changes, monitoring in production.
Tool integration with MCP
MCP servers over your ERP, CRM and internal systems, with scoped permissions, so assistants and agents can act safely.
03 — Commercial automation
Answer quote requests in 24 hours instead of 5 days, without hiring.
Our first product reads incoming requests, anonymizes them, extracts the technical data, pre-prices from your ERP and drafts the offer. Your sales engineer validates: nothing leaves without a human decision. Early-access pilots open in 2027, on the same base as lead qualification and a sales assistant.
A · Your inbox and ERP
Requests and references
Emails, PDFs, drawings, price lists and customer history, where they already live.
B · Cognitum layer
Read, anonymize, extract, pre-price
Structured extraction with a confidence score, pricing rules from your ERP, a draft offer ready for review.
C · Your people
Validation and measurement
The sales engineer decides. Response time, volume and conversion feed your dashboard.
04 — How we work
From a free workshop to a measured result.
01 · Free, 90 min
Discovery workshop
Bring your product, your process or twenty real documents. We show what is feasible today and what it would take.
02 · 5–8 days
Diagnostic
Architecture and process review, three to five quantified use cases, prioritized roadmap, compliance classification.
03 · 4–6 weeks
Pilot
Your real data, a success metric fixed in advance, an evaluation set you keep.
04 · 8–12 weeks
Build and integrate
Production integration, human validation, evaluation gates, dashboards, documentation.
05 · Subscription
Run and improve
Hosting, monitoring, model updates under evaluation, support and a quarterly review of gains.
05 — Why Cognitum
Production-grade, not demo-grade.
Twenty years of architecture, applied to AI
Senior engineers with more than twenty years of professional software practice. AI features land on sound architecture, inside the systems you already run.
Evaluation built into every project
Every system ships with an evaluation set and metrics agreed with the business, so model and prompt changes are measured, not hoped for.
Private AI when you need it
Open-weight models, hosting in France or on-site, and our own GPU infrastructure for pilots. Public cloud when it fits, private when it must.
A trained anonymization engine
Our personal-data detection models, trained on commercially licensed data, mask names, addresses and identifiers before any processing.
06 — Trust and security
Controls designed in from day one.
Hosted where the data is allowed to be
Public cloud, Azure France region, sovereign cloud or your own servers, set by the sensitivity of the data.
Anonymized before any external model
Personal data is detected and masked in the pipeline, so what reaches a model is what the workflow needs.
Human validation and audit trail
Approval steps on sensitive decisions, logs and evaluation as part of the operating model.
No lock-in
Open components and clear interfaces: models, hosting and frameworks can change without a rewrite.
Bring the product or the process. Leave with a measured plan.
A free 90-minute workshop on your product, your process or a sample of your documents: what AI can do today, what it would take, and what it would save.