The Scheduling Bottleneck
In the construction industry, generating an
accurate and optimized project schedule is
a notoriously tedious task. Project managers
historically spend countless hours manually
defining Work Breakdown Structures (WBS), estimating
task durations based on labor productivity rates, and
mapping out complex logical dependencies (such
as Finish-to-Start or Start-to-Start relationships).
This manual dependency mapping is prone to human
error, often resulting in delayed timelines,
misallocated resources, and cascading budget overruns.
The primary motivation for this research project
was to eliminate this bottleneck by introducing
an intuitive, conversational interface capable
of executing complex scheduling logic behind the scenes.
Methodology & Implementation
Our solution is built upon a sophisticated
multi-agent architecture orchestrated by LangGraph.
Instead of relying on a single monolithic LLM prompt,which
often hallucinates or loses context in complex
constraint-satisfaction problems,we distributed
the cognitive load across four specialized agents.
The Intent Agent first acts as
the conversational entry point, utilizing a local
Faster Whisper model to transcribe voice inputs in
real-time with sub-second latency, extracting
high-level project parameters. The Phase Agent
then acts as a retrieval engine, querying
a Neo4j Knowledge Graph to instantiate a
historically accurate WBS template tailored
to the specific building type.
Crucially, the system does not hallucinate task
durations. The Details Agent
dynamically identifies the mathematical variables
required for each task (e.g., floor area, concrete volume)
and prompts the user for specific dimensions.
Finally, the Scheduling Agent
bridges the gap between natural language
understanding and strict mathematical optimization.
It translates the gathered dependencies, formulas,
and user inputs into a constraint programming model
solved by Google OR-Tools. This ensures the final
schedule is not just a statistical guess, but
a mathematically proven, optimal timeline
respecting all physical construction constraints.
Conclusions & Future Scope
The resulting system successfully
demonstrates that Large Language
Models can be effectively bounded
and directed to perform rigorous
engineering planning when paired
with symbolic AI (Knowledge Graphs)
and strict mathematical solvers
(OR-Tools). User testing indicates
a near 90% reduction in the time
required to generate a preliminary
project baseline schedule, moving
from a multi-day manual process
to a 5-minute conversational
interaction. Furthermore, the
integration of a dual-model
real-time voice interface
(utilizing Whisper tiny.en
for live previews and medium.en
for high-accuracy finalization)
drastically improves system accessibility,
allowing site managers to initiate
planning without needing specialized
software training. Future iterations
of this architecture could be expanded
to include resource leveling
(optimizing for fixed crew sizes),
real-time cost estimation
dynamically linked to the schedule,
and continuous schedule updating
based on daily voice-reported site progress.
