Kanopi: an estimating engine with an accuracy claim that survives checking

What it is

Kanopi is LÏEF Development's estimating system. It reads a plan set, takes off the quantities, prices them from a sourced rate library and assembles a bid, with guards that refuse to ship a number the engine cannot stand behind.

The engine is older than the AI in it. Around 2014, at Tellus, Fowler encoded the per-trade metrics subcontractors price from into an in-house estimating system, so the firm could stop waiting three weeks on trade bids before telling a client what a project would cost; a detailed estimate took two days after that. In 2026 the same problem came back at LÏEF as slow outside takeoff vendors and scattered rate sheets, and he rebuilt that decision with current tools. The same estimate at LÏEF now takes about two hours.

As of September 25, 2026, Kanopi is in place with eight builders and developers pricing their own work across the country, and over $600 million in proposals has been bid through it nationally in 2026. The door is open to any architecture or build firm that gets hold of what it does. The work is ongoing.

What Common Ground did

• Built the pipeline from plans to a priced bid in April 2026. The first autonomous run priced a 6,919 SF premium spec home in 3 minutes 31 seconds, and the next three ran 2 to 4 minutes each at roughly a dollar of compute per run.
• Turned the first real mistake into a permanent gate. A May quick-service restaurant bid shipped with its markup stack compounding on itself; Fowler caught it, corrected it in the second version and made the check one that fails the build on its own instead of relying on a reviewer to notice.
• Refused to publish the flattering number. Fowler had a frontier model audit the engine for every place it was grading its own homework. It found that 75 of the 167 rates in the catalog had been transcribed from the anchor project's own budget, sorted all 96 formulas into calibrated, assumed and known wrong (40, 33 and 11), and found a six-layer verification stack with no production callers. He wired the gates in, added 24 hard assertions, and ran a leave-one-out test: each bid removed from the corpus, priced as if unseen and compared to what actually went out the door.
• Calibrated the formulas against a 41-bid corpus of LÏEF's own submitted bids, with a rule that no formula moves without corpus evidence. Of 54 formulas dispositioned, 36 could be calibrated and 18 could not, and those 18 carry their debt annotated in the engine. The pass surfaced two silent pricing defects, a premium finish priced at the mid tier and a life-safety trade that never emitted its line, both fixed before either reached a client bid.
• Ran three blind takeoffs in one day, each under an hour, for a building-materials manufacturer's VP of Operations who tested the engine unsolicited, one on a plan with no dimension strings. The takeoffs flagged a slab conflict inside his own set, and a blind second pass, with no access to the first, caught a wall-height error in the engine's own first pass that would have undersized the shell by roughly a third. He wanted to buy the method outright. LÏEF kept it, and the blind second pass became a standing rule on every takeoff.
• Ran a blind backcast on eleven past bids before anything went on a public page, and set the public rule: results only, method private.
• Rebuilt the front end in September 2026 to measure the drawing's own vector geometry, on the order of 49,000 paths on a single plan sheet, instead of a picture of it, with an overlay a person must inspect before any quantity is trusted and a ratio check beside it as a second control. The rebuilt engine also caught a real mid-project revision on another set: a 4.9 percent pixel change and a path count that moved from 36,646 to 76,994 between versions.
• Ran the same discipline on the design side, where a Revit skills library with a human in the loop turned a stalled as-built into a design-development set on a six-week agentic build and cut architectural production time 65 percent against a 50 percent target. A three-agent adversarial review caught a wrong setback before any person did, and when the same model was later found to carry a footprint problem traced to another project's plan set, that is exactly why the overlay rule and the blind pass exist.

Results

• Held-out accuracy, the only accuracy claim Kanopi makes: priced blind across three single-family bids, mean error of 18.6 percent, worst bid 31.9 percent (June 10, 2026). The figure is re-run as the corpus grows.
• An 11-bid blind backcast in which no miss traced to takeoff quantities, and a 121-unit hotel's building area landed within 1 percent.
• A full quantity takeoff went from about 40 hours to about 40 minutes, delivered in 24 to 48 hours with the blind pass and overlay included, at an internal cost under $350.
• The vector rebuild's first proof caught a five-times overcount on a 36,153 SF campus: 1,834 linear feet of interior partitions, not the 9,470 an earlier desktop estimate had carried. An independent blind pass landed within 2.4 percent, and eleven sheet-level contradictions went to the architect as requests for information.
• As of September 25, 2026, eight builders and developers price their own work through Kanopi, with over $600 million in proposals bid nationally in 2026.

Where we use it

We price LÏEF Development's own bids on Kanopi. It was built for LÏEF first, because outside takeoff vendors were slow and rate sheets were scattered, and its formulas are calibrated against bids LÏEF actually submitted, with no formula allowed to move without that evidence. Being the first customer is also how the first real mistake surfaced: a markup stack that compounded on itself went out on a live bid, was caught and corrected, and became the engine's first permanent gate. A builder who prices through Kanopi gets an engine whose mistakes we paid for first.

• LÏEF Development, wiki page
• LÏEF Midtown, 301 W Osborn data room
• Kanopi data room

External links

• Official website: Kanopi
• Public data room: Kanopi: an accuracy claim that survives being checked data room

Stories from this work

At a glance