ML Engineer (LLM Focus)

SWATX

Regulation Text Interpretation: Develop and refine LLM-based pipelines to parse and understand regulatory text (building codes, standards, etc.), extracting key conditions and requirements. For example, implement prompt-based or fine-tuned models (e.g. GPT-4) for one-shot or few-shot extraction of structured rules from text mdpi.com .

Rule Formalization: Convert interpreted regulations into structured representations (e.g. object-property-condition-value tuples or logical statements) that a compliance engine can use mdpi.com mdpi.com . This involves defining schema or labels for components of rules (such as what element is regulated, what property, the conditional operator, and the required value).

IFC Querying via NLP: Enable the system to query the BIM model s IFC metadata using natural language inputs or the extracted rules. Design methods for the LLM to map rule requirements to specific IFC entities and attributes. For instance, if a rule says the living room area must be 10 m , ensure the model can identify IfcSpace objects of type LivingRoom and retrieve their area property.

Collaboration on Knowledge Models: Work closely with the Computational and BIM engineers to align the LLM outputs with the BIM data schema and ontology. Ensure that terms in regulations correspond to the correct IFC classes or property names (e.g., ceiling height mapped to the right IFC attribute).

Model Integration: Integrate the LLM into the overall pipeline, potentially using retrieval augmented generation to fetch relevant model information. For example, use the LLM to generate queries or filters for the BIM model database given a regulatory clause.

Testing and Refinement: Evaluate the LLM s interpretations against sample regulations and expected outcomes. Refine prompts, training data, or model parameters to improve accuracy in capturing regulatory intent. Address edge cases where regulations are ambiguous or have nested conditions, ensuring the LLM can handle complex sentence structures mdpi.com .

Required Skills and Qualifications

Natural Language Processing (NLP): Strong background in NLP and language models, with experience in transformer-based LLMs (such as GPT, BERT, etc.). Understanding of prompt engineering and fine-tuning techniques for domain- specific language tasks.

Regulatory Domain Knowledge: Ability to grasp construction/building regulations or willingness to learn the terminology. Comfortable working with legal/technical text and extracting meaning.

Programming s ML Frameworks: Proficiency in Python and ML libraries (TensorFlow/PyTorch, HuggingFace Transformers). Ability to write scripts for data preprocessing (text cleaning, tokenization) and to implement inference pipelines.

Data Structures s Querying: Familiarity with JSON, XML, or other structured data formats for outputting rules. Basic understanding of databases or graph structures is a plus, to help formulate how queries to BIM data might be structured.

Analytical Skills: Excellent logical reasoning to design rule schemas and verify that the LLM outputs align with real-world logic. Attention to detail in handling units, thresholds, and conditional logic from regulations.

Collaboration: Good communication skills to work with engineers from other disciplines (geometry, backend) and ensure the language model s outputs are usable in their contexts.