Foreman AI Research
Empirical evaluations and technical research from the Foreman AI team. We measure what matters: recall fidelity, extraction accuracy, and real-world construction plan intelligence.
TRACE Series, Report 4 · Pilot
The heaviest pen cuts each elevation into a few dozen pieces, one compact model votes on them, and windows and doors snap to their own frames, so every edge is a line the architect drew. Good residential sheets tripled across four generations.
TRACE Series, Report 3
TRACE extended from walls to roofs: on 14 held-out roof plans, roof area lands within a 1.8% median error and 84.5% of ordinary roof area is recovered as correct roof planes, every edge snapped to the drawing.
TRACE Series, Report 2
Adapting TRACE's wall model to commercial plans: on sheets from unseen projects, held-out commercial wall F1 rises from 0.889 to 0.943, and the full pipeline finds 93% of wall length instead of 77%.
TRACE Series, Report 1
TRACE (Tokenized Representation of Architectural CAD Elements) turns a PDF's vector CAD elements into wall tokens and verifies them with compact learned models, lifting wall-geometry F1 from 0.777 to 0.912 on the same sheets and scorer.
Empirical Comparison of General-Purpose AI vs Domain-Tuned Plan Intelligence
This benchmark compares a state-of-the-art general-purpose LLM against Foreman AI when applied to real-world construction plan sets. Results reveal a consistent performance gap favoring domain-tuned intelligence.
Nearly one million pages of real construction plans for AI training and evaluation
This publication introduces a dataset of 7,550 construction plan sets totaling 954,892 pages—the first large-scale, vector-native corpus enabling meaningful model training and benchmarking on professional construction documents. Updated September 2026: the corpus now holds nearly 98,000 real engineered plans, about 3.6 million pages.
Stay informed
We release benchmarks and technical publications as we validate new capabilities. Join the list to receive updates.
