Project Guide
Each route starts with the strongest implemented evidence for that role, then links to supporting systems. The AEC RAG system is the sole flagship; embodied AI and computational design provide the other primary evidence areas.
AEC Code Compliance RAG is the deepest project: public-source ingestion, metadata-rich chunks, retrieval evaluation, citations, abstention, service contracts, and retained failures.
Continue through the Specification Assistant, Massing Explorer, and QS Workbench to see one AEC decision journey.
Public or clearly labeled synthetic inputs; document assistance and schematic decision support, not professional advice or project delivery.Construction Embodied Agent Simulator compares state and rendered-pixel policies, evaluates an unseen appearance shift, records action-filter interventions, and replays commands in planar physics.
The grid route planner and telemetry monitor provide narrower deterministic supporting experiments in the repository.
Simulator-rendered observations and rule-aware filtering; no physical-camera, foundation-VLA, ROS, mobile-robot, or hardware evidence.Constraint-Aware Massing Explorer separates hard checks from editable proxy objectives, generates reproducible geometry, compares against an equal-sized baseline, and exposes invalid options.
The QS geometry engine and tested massing-to-takeoff adapter show how selected geometry crosses a typed downstream boundary.
Rectangular proxy geometry and simplified environmental objectives; no CAD/BIM kernel, developed plan, calibrated simulation, or approvable design.AEC Code Compliance RAG publishes document-, chunk-, and page-level retrieval results, label provenance, uncertainty, ablations, citations, no-result behavior, and focused tests.
The Specification Assistant adds requirement state, approvals, conflict handling, audit events, and a manually labeled language stress set.
Local retrieval and deterministic extraction by default; no paid model dependency, autonomous web research, or production-agent claim.The embodied-agent comparison contains the strongest model-training evidence, including fixed holdouts, behavioral metrics, shift failures, and model cards.
Text classification and local model monitoring remain focused experiments rather than selected-work claims.
Compact local models and bounded service scaffolds; no large-scale training, cloud platform, sustained-load, or external deployment evidence.Evidence Pattern
Source code, local commands, interfaces, and generated outputs establish the implemented boundary.
Fixed datasets, baselines, metrics, and tests make the result inspectable and reproducible.
Retained errors, shift results, abstentions, and rejected handoffs expose where the system breaks.
Professional judgment, authority, deployment, and physical validation remain explicit review boundaries.