A5 — Applied Intelligence
Custom AI Systems
From experiment to production
Applied domain A5
A proof of concept is easy. Building an AI system that works reliably inside a real organisation is a different engineering problem. We connect models to enterprise data, applications, workflows and users.
AI that operates in the real world, not just one that demonstrates well.
- 01Data
- 02Training
- 03Evaluation
- 04Integration
- 05Deployment
- 06Monitoring
- 07Improvement
The real problems
The real problems → Our approach
- 01
The pilot that never ships
A demo built on a clean sample impresses, then stalls. Real data is messier, integration was never planned and nobody owns the model once the project team moves on.
Our approachOne gateway for every model
Every model call, whether to an on-premises model or an external API, passes through a single gateway that applies authentication, rate limits, personal-data filtering, logging and cost tracking.
- 02
Data that cannot leave
Customer records, contracts and internal documents are exactly what makes AI useful — and exactly what cannot be sent to an external service without controls.
Our approachRetrieval that respects access rights
Enterprise documents are indexed together with their permissions, so an assistant can draw only on what the person asking is allowed to read — and cites every source it uses.
- 03
Quality that is hard to measure
Language models can be fluent and wrong at the same time. Without an evaluation set and a scoring method, every change is a guess.
Our approachEvaluation before and after release
Each use case has a versioned evaluation set and a scoring method. Every model, prompt or retrieval change is tested offline first, then monitored in production for quality, drift, latency and cost.
- 04
Governance that arrives late
Risk, legal and security teams are asked to approve a system after it has been built. Without a model inventory, approvals and logs, the answer is often no — or a long delay.
Our approachGovernance built in
A model registry records versions, owners, approvals and intended use. Guardrails and human review are configured per use case, and user feedback flows back into the next improvement cycle.
The product for this domain
VARDA Aperture
Enterprise AI production platform
AI enters on your terms.
One controlled route for enterprise AI: model gateway, permission-aware retrieval, evaluation, guardrails, monitoring and an audited model registry.
- Model and LLM gateway
- Permission-aware retrieval
- Evaluation harness
- Guardrails and policies
Illustrative scenarios
Illustrative scenarios
Scenarios are illustrative; they do not describe specific clients.
- 01
A procedures assistant for the contact centre
ContextAt a regional electricity distribution company, contact-centre agents search hundreds of procedures and regulatory notices to answer customers, and answers vary from one agent to the next.
OutcomeAn on-premises assistant answers from current procedures only, cites the paragraph it relied on and hands the question to a supervisor when confidence is low. Answers become consistent, and outdated documents are flagged to their owners.
- 02
Contract review under confidentiality
ContextA holding company’s legal team reviews supplier contracts for non-standard clauses by hand. The documents are confidential and cannot be sent to an external AI service.
OutcomeAn open-weight model running on the group’s own GPUs flags deviations from the standard clause library and explains each one. Lawyers review the flagged clauses first, and every model output is logged.
- 03
From pilot to production
ContextA telecom operator’s churn-prediction pilot performed well on historical data but was never connected to the campaign system, and nobody knows whether it still holds.
OutcomeThe model is rebuilt on a reproducible pipeline, registered with an owner and approvals, and served to the campaign platform through an API. Drift monitoring and regular challenger comparisons keep its quality visible.
Compliance & governance
Compliance & governance
- Designed to support the risk-based obligations of the EU AI Act: each use case is assessed against the Act’s risk classes, and high-risk systems are documented with human oversight and event logging.
- Designed to support KVKK and GDPR requirements: personal data is filtered or masked before it reaches a model, and processing can be kept entirely within your environment.
- Designed to support ISO/IEC 27001-style controls, with role-based access to models and data and a log of every request.
- Designed to support model-governance expectations in regulated sectors: every model has an owner, documented validation and an approval trail.
Frequently asked questions
Frequently asked questions
01Why do AI pilots so often stall before production?
Usually for reasons unrelated to the model: data pipelines that only ever ran once, no integration into the systems where decisions are made, no owner, and no way to show risk and security teams how the system behaves. We plan those parts from the first day, so the pilot is the first release of the production system rather than a separate prototype.
02Can we use large language models without sending data outside?
Yes. Open-weight models can run on your own GPU servers, served with vLLM behind the same gateway as any external model. Where you do choose an external API, the gateway filters personal data and logs every request.
03How do you decide whether an AI system is good enough?
Against criteria agreed before it is built. Each use case gets an evaluation set drawn from real cases, including difficult ones, and a scoring method suited to it — accuracy, grounding in sources, refusal behaviour or a business outcome. A release that scores below the current version does not go live.
04Which models do you work with?
We are not tied to any single model. Depending on the task and your constraints, that may mean open-weight language models, commercial APIs or classical machine-learning models — often a combination. The gateway and evaluation harness let you switch models without rewriting the applications that use them.
Let’s design the right architecture for custom ai systems.
Taking AI from proof of concept to production: models connected to your data, applications and users, with evaluation, guardrails and monitoring.