Throughline — AI construction scheduling & project controls

A source-available, claims-grade construction scheduling and project-controls platform. A deterministic CPM/DCMA/EVM engine at the core; risk (SRA/Bayesian), resource optimization (RCPSP/Pareto), a reality-capture loop, and an LLM copilot layered around it as decision-support — never replacing the human or the deterministic schedule.

License: PolyForm Noncommercial 1.0.0 (see LICENSE/NOTICE) — free for research, teaching, and other noncommercial use; commercial use is by separate license — contact Alireza Shojaei shojaei@vt.edu. Status: pre-release; see What's validated below and docs/reports/RELEASE-READINESS.md. Real-site field validation is a post-launch activity.

What it does

• Deterministic core (the authority): CPM forward/backward pass with working calendars + constraints; DCMA-14 checks; Earned Value + Earned Schedule (SPI(t)); RCPSP/MRCPSP resource scheduling + multi-objective Pareto optioneering — all in-house, all tested against hand-worked oracles and public benchmarks.
• Risk & forecast: Schedule Risk Analysis (Monte-Carlo) + Bayesian duration priors → calibrated finish forecasts (P50/P80) with driver attribution.
• Reality-capture loop (M5): ingest progress (manual + 3D scan), map element↔activity↔zone, human-review the evidence, re-forecast. Scan→BIM registration + a 3D segmentation model feed it.
• Generative / AI moat (all gated): natural-language schedule authoring, BIM→schedule generation, and a copilot — every AI-proposed change is routed through the deterministic validator + a human approval gate (the LLM cannot write the authoritative schedule), and delay narratives cite evidence by a named method.
• Product: a FastAPI surface + a React/TS workspace (narrative / forecast / gantt / status / optioneering / baselines / reports).

What's validated (and on what data, with numbers)

This is an honest scope, not a sales sheet. Every number is recomputed deterministically by scripts/build_evidence.py (the committed docs/reports/EVIDENCE.md):

Capability	Dataset	Result
RCPSP optimality	PSPLIB J30 (committed instances)	solved to the published optimum (gap 0)
Forecast calibration	NYC capital projects (held-out)	P80 coverage 0.798 (Brier 0.25, CRPS 95 d)
EVM / Earned Schedule	Lipke hand-worked oracle	ES 4.5 (exact)
M5 re-forecast convergence	synthetic ground-truth overrun	P50 lands the known finish ±2 wd
Scan→BIM registration	synthetic corner cloud	fitness 1.0, inlier RMSE 6 mm
3D element segmentation	synthetic-from-BIM (9 IFCs)	held-out mIoU 0.2459 (dependency-light baseline)
3D element segmentation (real data)	real S3DIS, held-out Area-5 (Sonata/PTv3)	held-out mIoU 0.7384 — beats the synthetic baseline ~3×
3D element segmentation (real, shell construction)	real Rohbau3D, site-disjoint test (Concerto-large)	held-out mIoU 0.5840 (18-class) — the first published Rohbau3D segmentation
Scan↔BIM element confirmation (real data)	real SLABIM, session-disjoint (Concerto-large)	per-class IoU door 0.708 / column 0.897 vs the prior method FM-Fusion (0.443 / 0.316)
2D element segmentation (real data)	real HBD, held-out test (DINOv2-L)	held-out mIoU 0.5198 (9-class, image-level split; Window/Door detected)
2D defect segmentation (real data)	real S2DS, official test (DINOv3-L)	held-out mIoU 0.8436 — beats the cited SOTA (0.66) and the 16-method field (0.665) by +0.18
2D defect segmentation (real, bridge)	real dacl10k, public val (ensemble)	mIoU 0.4758 — beats the published baseline (0.414); honestly short of the closed-challenge SOTA (0.51)
Reality-capture quality loop	Demo Tower showcase (scripts/demo_capture_loop.py)	defect → human-gated rework → forecast-only re-time; baseline never touched

Honesty rules we hold ourselves to (see docs/honesty.md / model cards):

• We state exactly what is validated, on what data, with the achieved number — never a vendor-style "up to X% faster/savings" claim.
• The 3D-seg mIoU 0.2459 is a synthetic-only baseline. Real-data training on the GPU lane must beat it: a Sonata-pretrained PTv3 fine-tune on real S3DIS achieves 0.7384 mIoU on the held-out Area-5 (capture3d-real-v1, single-crop, no TTA), pinned in docs/models/REPRODUCIBILITY.json — the achieved held-out number, never aspirational. (The two use different vocabularies — synthetic 9-class IFC vs standard 13-class S3DIS — so are not class-for-class comparable; the contrast reflects the real-vs-synthetic leap on the standard benchmark.)
• AI is decision-support a human signs off, never an autonomous expert. Nothing AI-proposed is auto-applied to the schedule. A model-detected defect can only ever propose rework — it is always human-gated (no model confidence bypasses the gate) and only ever adjusts the forecast, never the approved baseline (ADR 0013).
• We lead with the per-class beat where the overall isn't a clean win: the SLABIM door/column IoU is the defensible scan↔BIM result; the overall mIoU is substrate-optimistic and not claimed as SOTA (model card).
• Delay narratives name their forensic method (e.g. AACE RP 29R-03 MIP) and cite the evidence; numbers must reconcile to engine facts or the narrative is rejected.
• Field generalization improves post-launch — the released models are honestly-scoped baselines.

Quickstart

pip install -e ".[dev]"
pytest -m "not benchmark" -q                 # the functional suite (Docker provides ephemeral Postgres)
python scripts/build_evidence.py             # regenerate docs/reports/EVIDENCE.md

One-command demo (API + UI + Postgres + a seeded sample): docker compose up (see docs/getting-started.md). GPU model training is a separate lane on a CUDA box — see GPU_TRAINING.md.

Architecture & decisions

• src/scheduling_platform/ — the engine + API; ui/ — the React/TS workspace.
• Design decisions are recorded as ADRs in docs/adr/ (baseline immutability, the LLM-never-writes rule, Redis-Streams queue, the PolyForm-Noncommercial license, …). Do not re-decide what's in an ADR.
• How we build (strict TDD; dual-subagent self-gating; per-component exit reports in docs/reports/) is in CONTRIBUTING.md.

Security & data

Automated supply-chain (pip-audit/pnpm-audit/SBOM/trivy), app-security (headers/CORS/limits/authz + bandit/ semgrep/CodeQL), and data-protection (tenant isolation, export + right-to-erasure) run in CI — see docs/security/ and SECURITY.md. Datasets are never redistributed here; provenance is in docs/datasets/SOURCES.md.

License

PolyForm Noncommercial License 1.0.0 — free for research, teaching, and other noncommercial use. Commercial use is not granted by this license; contact Alireza Shojaei shojaei@vt.edu for a commercial license. See LICENSE and NOTICE (third-party attributions, incl. the MPXJ LGPL arm's-length boundary).