ProjectVIEW AI Models: SLM + LLM

DANAOS ProjectVIEW AI combines domain-specific small language models (SLMs) with leading large language models (LLMs), and chooses the model per use case. SLMs handle pattern recognition and numerical analysis on project data. LLMs handle reasoning, language and multi-step planning. Both work on ERP context that the system retrieves at the moment of the question. As a result, no model needs retraining on your data.

For construction and other project-based industries, this mix of SLM and LLM matches the work: precise numbers on one side, explanation and judgement on the other.

A large language model is a general model trained on broad text. It is strong at understanding questions, writing, summarising and planning a sequence of steps. A small language model is a smaller model focused on a narrower domain or task. It is typically faster and cheaper to run, and teams can tune it to recognise patterns in specific kinds of data.

Neither is better in general. The question is which one fits the task.

Project data in construction, shipbuilding and other project-based industries mostly holds numbers in a fixed format: quantities, unit rates, commitments, certifications, retention, fuel use. It also comes with its own language: BoQ, WBS, Cost Codes, variations, claims.

• Numbers need consistent, repeatable analysis. That is SLM work.
• A director needs to know why a project is over budget, in plain words. That is LLM work.
• The ERP holds the facts. That is deterministic work.

Deterministic ERP + probabilistic AI. ProjectVIEW ERP keeps the record exact. The models add detection, explanation and forecasts, and the user stays responsible for decisions.

The table shows how tasks typically split between model types. However, we select the actual model for each use case per deployment.

Take a director’s question: “Why is Project Alpha over budget?”

1. Retrieve. ProjectVIEW AI retrieves Project Alpha’s budget, committed cost, actual cost, progress and open variations, within the director’s permissions.
2. Analyse. Numerical analysis identifies which WBS elements and Cost Codes drive the deviation and how far each one is off.
3. Explain. An LLM turns the result into a short explanation in plain language, with the figures and the source lines.
4. Recommend. Where relevant, it suggests what to review next: an unapproved variation, a supplier above budget rate, a delayed certification.

The director gets an answer that they can check line by line in ProjectVIEW cost control. The decision stays with the director.

ProjectVIEW AI is model-flexible: we choose the model per use case rather than fixing one for the whole product. A numerical forecast and an executive summary do not need the same model.

Being model-flexible also means the AI layer can adopt better models as they appear, without changing the structure of your data in ProjectVIEW ERP. The structure in the ERP is the stable part; the models around it can change. Client-specific SLMs are available in ProjectVIEW AI v12.

ProjectVIEW AI retrieves ERP context at inference time. When a user asks a question, the system retrieves the relevant records from ProjectVIEW ERP. Then it passes them to the model with the question. Therefore, the model does not need retraining on your data.

1. Current answers. The model reads today’s commitments, not last quarter’s training snapshot.
2. Permissions respected. Retrieval follows the user’s role, so the model only sees what the user may see.
3. No data baked into a model. Also, your commercial data never becomes part of the model weights.

This is how ProjectVIEW AI learns your business: through your structured ERP data, not through guesswork. Read more in how ProjectVIEW AI works.

No. Data never goes to public consumer AI services, and ProjectVIEW AI runs in a private environment. In addition, on Microsoft Azure, Microsoft does not use prompts for model training. See ProjectVIEW AI data security and governance and the background article on private enterprise AI for construction.