EM
E. Murillo Sierra
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
This thesis examines how multi-agent systems based on large language models can close persistent design feedback gaps in Target Value Design within Integrated Project Delivery. In current practice, estimates arrive after key choices, non-cost value is rarely tracked, and critical data remains siloed. These conditions push teams toward reactive budgeting rather than value-oriented design. Employing a Design Science Research approach embedded in Stanford University’s AEC Global Teamwork project as a case study, the work develops and evaluates an Agentic Target Value Design workflow that tightens loops between quantities, costs, and value targets while keeping processes transparent across disciplines.
The system, built on the open-source AG2 framework, organizes GPT-4.1 agents in a planner, manager, and worker hierarchy that coordinates Python tools for model-based quantity takeoff, cost estimation, and TVD checks over IFC, CSV, and XLSX inputs. A retrieval-augmented generation subsystem links to a PDF cost repository. Agents share context through controlled handovers and restricted tool calls, and a nested chat agent stores iteration summaries in a vector database for natural language querying.
In a workshop with seven AEC Global Teamwork participants (architects, engineers, life cycle financial managers, and construction managers), users reported shorter wait times for design feedback, greater transparency into estimation steps, and improved grasp of cost and value trade-offs during design. The conversational interface supported distributed collaboration and encouraged more frequent TVD iterations while preserving trust. Limitations include a small number of design cycles and information loss for some IFC elements. Future work should broaden testing and integrate life cycle assessment and
richer knowledge structures such as GraphRAG and knowledge graphs. Overall, the study shows that Agentic Target Value Design can shift TVD from delayed cost control to an interactive, value-oriented dialogue, advancing both IPD practice and agentic AI in AEC. ...
The system, built on the open-source AG2 framework, organizes GPT-4.1 agents in a planner, manager, and worker hierarchy that coordinates Python tools for model-based quantity takeoff, cost estimation, and TVD checks over IFC, CSV, and XLSX inputs. A retrieval-augmented generation subsystem links to a PDF cost repository. Agents share context through controlled handovers and restricted tool calls, and a nested chat agent stores iteration summaries in a vector database for natural language querying.
In a workshop with seven AEC Global Teamwork participants (architects, engineers, life cycle financial managers, and construction managers), users reported shorter wait times for design feedback, greater transparency into estimation steps, and improved grasp of cost and value trade-offs during design. The conversational interface supported distributed collaboration and encouraged more frequent TVD iterations while preserving trust. Limitations include a small number of design cycles and information loss for some IFC elements. Future work should broaden testing and integrate life cycle assessment and
richer knowledge structures such as GraphRAG and knowledge graphs. Overall, the study shows that Agentic Target Value Design can shift TVD from delayed cost control to an interactive, value-oriented dialogue, advancing both IPD practice and agentic AI in AEC. ...
This thesis examines how multi-agent systems based on large language models can close persistent design feedback gaps in Target Value Design within Integrated Project Delivery. In current practice, estimates arrive after key choices, non-cost value is rarely tracked, and critical data remains siloed. These conditions push teams toward reactive budgeting rather than value-oriented design. Employing a Design Science Research approach embedded in Stanford University’s AEC Global Teamwork project as a case study, the work develops and evaluates an Agentic Target Value Design workflow that tightens loops between quantities, costs, and value targets while keeping processes transparent across disciplines.
The system, built on the open-source AG2 framework, organizes GPT-4.1 agents in a planner, manager, and worker hierarchy that coordinates Python tools for model-based quantity takeoff, cost estimation, and TVD checks over IFC, CSV, and XLSX inputs. A retrieval-augmented generation subsystem links to a PDF cost repository. Agents share context through controlled handovers and restricted tool calls, and a nested chat agent stores iteration summaries in a vector database for natural language querying.
In a workshop with seven AEC Global Teamwork participants (architects, engineers, life cycle financial managers, and construction managers), users reported shorter wait times for design feedback, greater transparency into estimation steps, and improved grasp of cost and value trade-offs during design. The conversational interface supported distributed collaboration and encouraged more frequent TVD iterations while preserving trust. Limitations include a small number of design cycles and information loss for some IFC elements. Future work should broaden testing and integrate life cycle assessment and
richer knowledge structures such as GraphRAG and knowledge graphs. Overall, the study shows that Agentic Target Value Design can shift TVD from delayed cost control to an interactive, value-oriented dialogue, advancing both IPD practice and agentic AI in AEC.
The system, built on the open-source AG2 framework, organizes GPT-4.1 agents in a planner, manager, and worker hierarchy that coordinates Python tools for model-based quantity takeoff, cost estimation, and TVD checks over IFC, CSV, and XLSX inputs. A retrieval-augmented generation subsystem links to a PDF cost repository. Agents share context through controlled handovers and restricted tool calls, and a nested chat agent stores iteration summaries in a vector database for natural language querying.
In a workshop with seven AEC Global Teamwork participants (architects, engineers, life cycle financial managers, and construction managers), users reported shorter wait times for design feedback, greater transparency into estimation steps, and improved grasp of cost and value trade-offs during design. The conversational interface supported distributed collaboration and encouraged more frequent TVD iterations while preserving trust. Limitations include a small number of design cycles and information loss for some IFC elements. Future work should broaden testing and integrate life cycle assessment and
richer knowledge structures such as GraphRAG and knowledge graphs. Overall, the study shows that Agentic Target Value Design can shift TVD from delayed cost control to an interactive, value-oriented dialogue, advancing both IPD practice and agentic AI in AEC.