From the PMC teamDevelopment projects start in 2027. Plan yours now to secure an early start.

AI development

AI that does the work, not only the talking.

We build assistants and agents on your own data. They read your documents, look things up in your systems through MCP, prepare the next step and act, and a person approves what matters. On Claude, OpenAI, Mistral or open models on your own servers.

  • Fast to deliver
  • Built to last
  • Secure from day one
11+years of building software for finance, IT and project teamsAbout the team
AI development

Our AI development services

From choosing the first use case to running the assistant every day. Pick one step or let us carry the whole way.

AI strategy and use cases

We pick the processes where AI pays off first and agree how success is measured.

Discuss this service

What we cover: Any model, any platform: OpenAI, Azure OpenAI, Anthropic, Google, Mistral or open models on your own servers. It runs where you decide: in your cloud, on-premises or on servers we run for you.

The team behind PMC, trusted by
  • BNP Paribas
  • Allianz
  • Hexagon
  • Hewlett Packard Enterprise
  • Enapter
Solutions

AI solutions we build

  • Invoice and contract checks

    Every incoming invoice is compared with the order and the contract, and only the differences reach a person.

  • Service desk assistant

    First answers for employees or customers from your knowledge base, and a clean hand-over to the team when it cannot help.

  • Tenders and proposals

    A first draft from your past proposals, product sheets and the tender text, ready for the expert to finish.

  • A copilot for controllers

    Questions such as "which projects will overrun this quarter?" answered in plain language from your budget and ERP data.

Your teams answer the same questions every week?

Show us three of them. We tell you within a day whether an assistant can answer them, and from which sources.

Plan your assistant
A worked example

How a project like this runs.

Example

Checking supplier invoices against orders and contracts

The situation

A finance team receives several hundred supplier invoices a month and checks each one against the purchase order and the framework contract by hand.

What we do
  1. We take a sample of past invoices with known outcomes as the test set.

  2. Document AI reads each invoice; an agent looks up the order in the ERP through MCP and the prices in the contract.

  3. Matching invoices are marked ready to book; differences go to the accountant with the reason.

  4. Nothing is booked without a yes, and every step is logged.

The result

The team checks the exceptions instead of every invoice, and every decision can be traced.

An illustrative example of the way we work, not a named customer.

AI inside

What our assistants and agents can do

Combined per use case, on the model and the servers you choose.

  • Answers with sources
  • Semantic search
  • Document reading and OCR
  • Actions through MCP tools
  • Drafts for review
  • Speech and voice input
  • Classification and routing
  • Approval before every action
  • Cost and quality dashboards
Why it pays

Why build AI on your own data

01

Answers in seconds, with the source

Today
Experts search folders, SharePoint and old e-mails for the same answers every week.
With it in place
An assistant answers from your documents in seconds and links the source of every statement.
Ready for production

Not a demo. A system you can rely on.

What separates an AI pilot from a tool your teams use every day.

Measured before it goes live
Every assistant is tested against a set of real questions with known answers, and tested again after every change.
A person approves what matters
Agents prepare; bookings, e-mails to customers and changes to master data wait for a yes.
Every step logged
Which question, which documents, which tool call and which answer: traceable for audit and improvement.
Costs under control
We measure cost per request and pick the smallest model that does the job well.
What you receive

Everything you need to own the result.

  • The assistant or agent, running in your cloud, on your servers or ours
  • The test set of real questions and the result of every evaluation
  • Prompts, tools and configuration in your repository
  • A log of every question, source and tool call
  • A cost report per request and per month
  • Documentation and training for the people who use and run it

Is it right for you?

A good fit when
  • A task that repeats every week and follows recognisable rules
  • Knowledge that exists in documents or systems, even if it is scattered
  • A person who owns the process and can judge a good answer
Not the right fit when
  • Decisions that must be taken without any human check
  • Processes whose underlying data does not exist yet
Speed, quality and security

Fast does not have to mean fragile.

Every project is held to the same standard, whatever the service. It is the standard we build and run PMC to every day.

  • Fast to deliver

    Results in weeks, not quarters.

    • A working version at the end of every short cycle
    • Releases through an automated pipeline, not by hand
    • Small changes shipped often, so value arrives early
  • Built to last

    Software your team can still understand and change years from now.

    • Every change built and tested automatically before it is merged
    • Unit, API and end-to-end tests in a real browser
    • Documentation and a clean handover, or we run it for you
  • Secure from day one

    Security is part of the build, not a check at the end.

    • OWASP Top 10 and BSI IT-Grundschutz as the reference
    • Dependencies and images scanned for known vulnerabilities in every build
    • Access by role, changes logged, secrets kept out of the code
How we work

From the first question to an assistant your team relies on.

  1. Pick one process

    We choose one process that costs time every week, such as invoice checks, and agree how success is measured.

  2. Build on your data

    A first version works on your own documents and systems, in your cloud or on your own servers.

  3. Measure before go-live

    The assistant is tested against real questions with known answers until it meets the agreed bar.

  4. Run and improve

    In daily use every step is logged, cost per request is tracked, and every change is tested again.

An AI pilot that never made it into daily work?

We measure it against real questions, fix what fails and take it into production, or tell you honestly why it should stop.

Review your pilot
Ways to work with us

Choose how much of the work you hand over.

From one clearly scoped project to a team inside your business and the long-term running of what we built. Every model has one accountable partner: Virtual Eye.

  • For a defined result

    Fixed-scope project

    You know what you get before the build starts: a concept, a prototype and a fixed scope come first.

    • A workshop on your process and the result you need
    • Concept, prototype and fixed scope before production code
    • A working version at the end of every short cycle
  • When the problem is not yet clear

    Forward deployed team

    A small Virtual Eye team works inside your business, learns how work and money really move, and delivers the solution with your people.

    • One lead for management, plus the engineers who build
    • Outcome and scope agreed in writing
    • A demo every week
  • For software you already run

    Takeover and modernisation

    We take over an existing application, make it safe to change, then develop it further step by step.

    • What exists is documented and tested first
    • Stabilised before anything new is added
    • Further development in small, tested releases
  • After go-live

    Run and evolve

    We host, monitor, update and support your applications on our servers, in your cloud or on-premises.

    • Monitoring and alerts that go straight to our team
    • Backups, updates and security patches
    • Support and further development from the team that built it
Customers

What customers say about working with us.

Quotes about the Virtual Eye team from PMC projects: the same people who would work on yours.

“Well thought-out concepts, tailored precisely to us. Very competent and always friendly, even during the stressful implementation phase.”
Iris HummelManager Infrastructure Services Sales DACH, Hewlett Packard EnterpriseTranslated from German
“PMC provides valuable support in day-to-day business and the team is very reliable, serious and friendly!”
Hao LiCentral Project Office PMO, Allianz Global Corporate & Specialty SE
“As a customer from day one, we have found reliable, fitting solutions with Virtual Eye throughout the years. Above all, the fast and flexible service.”
Sebastian StepferExecutive Director Public Safety EMEA Region, HexagonTranslated from German
Tech stack

The tools we use for this.

We choose them for your systems and your team, and we document every choice.

Assistants, agents, document AI and machine learning, on hosted models or your own servers.

Models and platforms

  • Claude
  • OpenAI
  • Azure OpenAI
  • Mistral AI
  • Google Gemini
  • Meta Llama
  • Ollama
  • vLLM
  • Hugging Face
  • Amazon Bedrock
  • AWS SageMaker
  • Azure Machine Learning
  • Google Vertex AI

Frameworks and libraries

  • Model Context Protocol
  • LangChain
  • LangGraph
  • LlamaIndex
  • PyTorch
  • TensorFlow
  • Keras
  • scikit-learn
  • XGBoost
  • CatBoost
  • LightGBM
  • OpenCV
  • OCR
  • NumPy
  • pandas
  • Jupyter

MLOps and vector databases

  • MLflow
  • Kubeflow
  • Pinecone
  • Weaviate
  • Milvus
  • Qdrant
  • ChromaDB
Your technology is not listed?

Tell us what you run. If we have not worked with it yet, we say so, and tell you how we would approach it.

Talk to our engineersSee all technologies
Industries we serve

AI for the industries we know

Years of PMC have taught us how these industries approve, buy, report and get audited.

Banks

Assistants that answer from policies and procedures, with every source traceable for audit.

  • Policy and procedure assistants
  • Invoice checks against contracts
  • Answers limited to each user's rights
Questions

Questions teams ask before they start.

Does our data leave the company?

Only if you decide so. Assistants can use Claude, OpenAI or Mistral under your own contract, or open models on your own servers, so documents never leave your network.

How do you prevent wrong answers?

Assistants answer from your sources and show them. Before going live they are measured against real questions with known answers, and uncertain cases go to a person.

Where should we start with AI?

With one process that costs time every week, such as checking invoices or answering recurring questions. A first version on your own data shows early whether it pays off.

Which models do you use?

The one that fits the task, the cost and where your data may go: Claude, OpenAI, Azure OpenAI, Google, Mistral or open models on your own servers. We measure cost per request and pick the smallest model that does the job well.

Can an agent act in our systems on its own?

It prepares the work and uses your systems through MCP with the permissions you set. Bookings, e-mails to customers and changes to master data wait for a person to approve them.

Tell us what you want to build.

Describe the process, the systems involved and when it should go live. We reply with a first assessment and the next step.

Or write to us directly: hello@virtual-eye.com

What happens next?

  1. You write to us

    What you want to build, the systems involved and when it should go live.

  2. A first assessment

    We reply with a first assessment and the next step.

  3. A workshop

    On your process and the result you need, before any technology is chosen.

  4. A fixed scope

    A concept, a prototype and a fixed scope before the first line of production code.