Measure before promising
A test set built from your own material and an agreed success metric. We compare models on your task before choosing.
Back to homeThis is how
Technical brief
This is how
we build it.
For CIOs, CTOs and CISOs: architecture, tools, deployment models, engineering practice and security.
T1 — Architecture
What an AI system
we build looks like.
Not a chatbot bolted onto the organization. Five layers, each with an owner, permissions and a log, all inside your network. Select a layer to see what happens in it.
YOUR PERIMETER · your organizational network
- 01 · SOURCESSources
- Documents and email
- ERP / CRM
- BI and databases
- Cameras and sensors
Data stays where it is: documents, core systems, email and files, cameras and sensors. We don’t copy your organization into someone else’s cloud.
- 02 · CONNECT & DATAConnect and data
- Connectors and APIs
- Cleaning and normalization
- Vector index
- Per-user permissions
We connect to existing interfaces, clean and normalize the data, and build a vector index that respects each user’s permissions. This is the layer most AI projects skip, and where they fail.
- 03 · MODEL GATEWAYModel gateway
- Claude
- GPT
- Gemini
- Local LLM
One layer in front of every model: Claude, GPT, Gemini or an open-weight model running in your environment. Routing by data sensitivity, task complexity and cost. Swap models without rebuilding the system.
- 04 · AGENTSAgents and workflows
- Agents
- Tools and actions
- Human approval
Agents that do work inside a defined process: retrieve, analyze, draft and update systems. Any action with consequences goes through human approval.
- 05 · INTERFACESInterfaces
- Teams and email
- Your existing system
- Dedicated interface
People work where they already work: Teams, email or the existing system. A dedicated interface only when it is really needed.
CONTROL PLANE
- Audit log
- Evals
- Cost monitoring
- DLP
- RBAC
Controls that run across every layer: a log for every answer, continuous quality metrics, cost monitoring, data-loss prevention and role-based access.
T2 — Technology
The tools
we work with.
We choose tools by requirement, data sensitivity and cost, and justify every choice in the architecture document.
LLMS
Language models
- Claude
- GPT
- Gemini
- Qwen
- Llama
AI ENGINEERING
AI engineering
- RAG
- Vector search
- Agentic
- Multi-agent
- Evals
MULTIMODAL
Vision, speech and multimodal
- YOLO
- VLM
- NLP
- Speech
CLOUD & INFRA
Cloud and infrastructure
- AWS
- GCP
- Vertex AI
- Docker
- On-prem
BUILD & DATA
Build and data
- Python
- FastAPI
- Node.js
- React
- PostgreSQL
SYSTEMS
Systems and automation
- REST APIs
- n8n
- ERP
- CRM
- BI
EDGE
Edge computing
- NVIDIA Jetson
- Hailo
DEV TOOLS
Developer tools
- Claude Code
- Cursor
- GitHub Copilot
T3 — Deployment
Where does it run?
Wherever it needs to.
Four deployment models, from cloud to an air-gapped network. Chosen by data classification.
The strongest models, on enterprise terms.
Commercial models through an enterprise account: no training on your data, a defined data region and log retention under your policy.
- Models
- Claude · GPT · Gemini
- Data
- Encrypted in transit to the provider, no training
- Fits
- Unclassified business data, where answer quality decides
Everything in your cloud account.
The system, data and indexes live in your organization’s AWS or GCP account, behind your VPN. Models are consumed through cloud services you have already approved.
- Models
- Vertex AI · Amazon Bedrock
- Data
- Stays in your cloud account
- Fits
- Organizations with an existing cloud policy
Your servers. No request leaves.
Open-weight models run on your GPUs, next to the data. For when data cannot leave the building.
- Models
- Open-weight models (Qwen, Llama and others)
- Data
- Never leaves the network
- Fits
- Legal, finance, healthcare
Even with no internet at all.
A fully disconnected network. Models, indexes and updates come in through a controlled, documented process.
- Models
- Pre-vetted open-weight models
- Data
- Never leaves, not even for updates
- Fits
- Defense and classified data
The further you move from the cloud, the smaller the models. We measure the gap on your tasks before choosing.
T4 — Engineering practice
How we engineer
AI for production.
The difference between a good demo and a system that still works in year two.
See what happens in production
Every answer is logged with its source, model, latency and cost. When something shifts, you see it.
Predictable cost
Routing between models by complexity and sensitivity, with a monthly cost ceiling set in advance.
The agent proposes, a person decides
Every action with consequences needs human approval, with a record of who approved what.
Only what you’re allowed to see
The system inherits permissions from existing systems. Nobody gets an answer from a document they aren’t allowed to open.
No black box
Code, documentation and training stay with you. You can continue without us.
SECURITY BY DESIGN
Security is part
of the design.
Your security team reviews and approves the design before building starts. The system is built on your infrastructure, and the deployment model is chosen together based on data classification.
- Need-to-know access and role-based permissions
- No model training on your data
- Encryption in transit and at rest
- An audit log for every action
- Aligned with Amendment 13 and GDPR
- Documentation that stays with you
Let’s find your advantage
What do you want
to move forward?
30 minutes on your processes, systems and blockers. We’ll look at what is worth building, what to buy and what not to do at all. No commitment.
Book an intro call