The Seam — Daily — July 29, 2026
THE SEAM · DAILY
Daily Briefing
Wednesday · July 29, 2026
Where AI meets the built environment
For a month the throughline here has been that capability is no longer the question — control is. This week the abstraction turned concrete. Walk the lifecycle and you can watch each player reach for a different layer to own: a town's plan review, an authoring tool's grip on your geometry, the training data under a document AI, the intelligence bolted to a rented machine, the software brain inside a building. Five segments, one question underneath all of them: whose layer is it?
1 · Pre-Development · Feasibility
A town outsourced its permit backlog to an AI — and got both speed and overreach
East Hampton, New York contracted four outside firms to clear a Building Department backlog running past a year. One of them, GreenLite, runs its own AI-assisted plan-review software, now live under a municipal contract. Builder Michael Forst says reviewed plans come back within about 10 days, versus the town's usual six-months-plus, and he endorses it. But real-estate attorney Jon Tarbet describes the failure mode: the model flagged electrical circuit labeling before an electrician was even hired, and raised demands that weren't on the town's 100-point checklist. GreenLite says it doesn't train on the plans it reviews, only on feedback patterns over time. The speed is real; so is the overreach — and "learns from feedback" is a data-governance claim a municipality should be able to audit, not just take on faith. 27East · Jul 22 →
2 · Pre-Design
Autodesk shows a model that reasons about geometry — and asks for more of your trust
In a deep interview with AEC Magazine's Martyn Day, Autodesk detailed Neural CAD: an in-house foundation model trained on professional CAD geometry, which reasons directly about B-rep shapes, topology, and engineering relationships rather than writing API commands. Its first AEC use inside Forma is conceptual massing and layout exploration — output as logical parametric systems (walls, doors, columns) with semantics, not just meshes. Autodesk paired it with "project intelligence," a persistent knowledge layer spanning the lifecycle, and, Day notes, quietly walked back its spring claim that Forma replaces Revit. This is the most credible pre-design AI we've seen in weeks — and the first real signal in a segment that's been an honest gap. The ask underneath it, though, is to entrust still more of your IP to one vendor's cloud. Read that part twice. AEC Magazine · Jul 21 →
3 · Design · Documentation
A benchmark says the generic document AI in your stack degrades on engineering drawings
A study accepted to the 2026 European Conference on Computing in Construction built the first AEC-specific drawing-layout dataset and benchmarked five architectures for pulling structured information off engineering sheets. A purpose-built detector, RF-DETR, hit 0.949 mAP for layout detection, and a vision-language model, Qwen3-VL, led information extraction at 0.911 F1. The sharper finding is the inverse: models pre-trained on general document sets suffer "domain interference," actively degrading on drawings because they're tuned for text-centric pages. Translation for anyone weighing an off-the-shelf "AI document reader" for your sheet sets: general-purpose tools aren't merely weaker on AEC drawings — they can be worse than doing nothing. The model layer has to be domain-built. arXiv / EC3 2026 · Jul 21 →
4 · Construction · Robotics
Gritt's bet: don't sell contractors new robots — bolt AI onto the machines they already rent
Carnegie Mellon roboticists brought Gritt out of stealth with $32.4M (Series A led by Obvious Ventures). The twist isn't a proprietary machine — it's software plus arms that mount on rented skid steers and off-the-shelf Kawasaki units. On utility-scale solar, they claim an 8-person crew's output jumps from ~800 panels a day to 3,000–4,000, with zero panel breakages across tens of thousands of autonomous placements; a rebar-tying skill was trained in a day. TechCrunch reports three of the top-10 U.S. power builders as customers. The honest caveat: only two systems are deployed against a 48-system, 2.8-GW pipeline, so that's a contracted claim, not delivered capacity. But "own the intelligence, rent the iron" is the most capital-efficient answer yet to the labor-shortage robotics pitch. TechCrunch · Jul 21 →
5 · Post-Occupancy · Building Ops
Carrier didn't build the AI brain for its buildings — it bought it
Carrier closed a full acquisition of 75F, the Minneapolis maker of cloud-native, AI-driven building automation (its Saffron AI platform, 1,800+ installations across nine countries). Carrier was already a minority investor from 75F's 2025 Series B; buying it outright converts a stake into ownership of the software layer that turns sensor data into automated decisions — which it will fold into WebCTRL controls, its Abound analytics, Nlyte, and its QuantumLeap data-center cooling suite. Unite.AI's read is the useful one: the "autonomous, self-optimizing building" language is a strategy slide, not a shipping product with a price or a date. For operators the signal isn't the demo — it's consolidation. The intelligence layer for building ops is being bought up by the incumbents who already own the hardware. Unite.AI · Jul 23 →
General AI — In Brief
Google shipped its Flash workhorse tier on July 21 — Gemini 3.6 Flash (−17% output tokens, better coding and multimodal), 3.5 Flash-Lite (~350 tokens/sec), and a government-restricted 3.5 Flash Cyber — but the 2M-token Gemini 3.5 Pro this digest has tracked since early July is still in partner testing with no date, and Gemini 4 is reportedly already training. AEC read: the models firms will actually run on the document grind keep getting cheaper and faster, while the "query your whole BIM model" flagship keeps slipping. Resolves our open watch with a real outcome — Flash first, Pro later. TechCrunch · Jul 21 →
Watching
Whether Gritt scales from two systems to its promised 48 in six months — the test of retrofit robotics as a category rather than a demo. The Next Web →
The BIM-access sovereignty move: the Open Design Alliance — a non-profit that cannot be acquired — opened its DWG/IFC/Revit SDKs as self-hosted MCP servers, and Revizto shipped a real-time Object Properties API, both this week. The question for firms: self-host access to your own model data, or rent it through a vendor's gateway? AEC Magazine →
The Seam · AI + AEC · A personal project by Bruce Lanier
Daily · July 29, 2026 · Verified-first, non-advertorial
Don't miss what's next. Subscribe to Bruce Lanier: