Work record / Estimating / Kanopi

Kanopi: an estimating engine whose accuracy claim survives checking

An estimate is only as good as the number you can defend. I built Kanopi so it refuses to ship a figure it cannot stand behind, and so the only accuracy we publish is the kind tested blind.

The seat

I am Principal at LÏEF Development and I hold the estimating seat. Kanopi is the system I rebuilt in 2026 to settle a problem I had met before, and the work is still going.

Sector	Construction and real estate development
Type of work	Applied AI system build and validation
Years	2026 to present
How long	Ongoing since 2026
Through	Kanopi, LÏEF Development's estimating system
Credits	Common Ground Jesse Fowler, Principal, LÏEF Development, in the estimating seat

What I was brought in to decide

Kanopi is LÏEF's estimating system. It reads a plan set, pulls the quantities, prices them against a sourced rate library and assembles the bid. Guards inside it stop any number the engine cannot defend from going out.

The engine is older than the AI inside it. Around 2014, at Tellus, I encoded the per-trade metrics that subcontractors price from into an in-house estimating system. Before that, the firm sat for three weeks waiting on trade bids before it could tell a client the cost of a project. After it, a detailed estimate took two days. In 2026 the same problem returned at LÏEF, now as slow outside takeoff vendors and scattered rate sheets. I rebuilt the decision with current tools, and today that estimate at LÏEF takes about two hours.

As of September 25, 2026, eight builders and developers across the country price their own work in Kanopi, and it has carried over $600 million of proposals bid nationally this year. Any architecture or build firm that learns what it does can come in. The work continues.

What we did

In April 2026 I built the pipeline that takes plans to a priced bid. On the first autonomous run, a premium spec home of 6,919 SF was priced in 3 minutes 31 seconds. Three more runs followed, 2 to 4 minutes apiece, with compute coming to roughly a dollar a run.
The first real mistake became a permanent gate. A May quick-service restaurant bid went out with its markup stack compounding on itself. I caught it and corrected it in the second version. Then I made the check one that fails the build by itself, so no one depends on a reviewer noticing.
I would not publish the flattering number. I asked a frontier model to audit the engine and hunt every spot where it graded its own homework. Of the 167 rates in the catalog, it found that 75 had been copied from the anchor project's own budget. It sorted all 96 formulas into calibrated, assumed and known wrong, which came to 40, 33 and 11. It also found a six-layer verification stack that nothing in production called. I wired the gates in and added 24 hard assertions. Then I ran a leave-one-out test: remove each bid from the corpus, price it as if unseen, and compare the result with what went out the door.
We calibrated the formulas against a corpus of 41 bids that LÏEF had submitted itself, under a rule that no formula changes without corpus evidence. We dispositioned 54 formulas: 36 could be calibrated, 18 could not, and the engine carries the debt for those 18 as annotations. The pass exposed two silent pricing defects. A premium finish had been priced at the mid tier, and a life-safety trade never emitted its line at all. Both were fixed before either reached a client bid.
In one day we ran three blind takeoffs, each under an hour, for the VP of Operations at a building-materials manufacturer who had tested the engine on his own initiative. One plan came without dimension strings. In his own set, the takeoffs flagged a conflict in the slab. Then a blind second pass, run without sight of the first, found a wall-height error in the engine's own first pass. It would have left the shell roughly a third undersized. The VP wanted to buy the method outright, but LÏEF kept it. After that, every takeoff gets a blind second pass as a standing rule.
Before anything went on a public page, we ran a blind backcast on eleven past bids and set the public rule: results only, method private.
In September 2026 we rebuilt the front end so it measures the drawing's own vector geometry rather than a picture of it. On a single plan sheet that is on the order of 49,000 paths. A person must inspect an overlay before any quantity is trusted, and a ratio check sits beside it as a second control. On another set, the rebuilt engine caught a real revision made mid-project, shown by a 4.9 percent pixel change and by the path count moving from 36,646 to 76,994 between versions.
We applied the same discipline on the design side. A Revit skills library with a human in the loop moved a stalled as-built to a design-development set. It was a six-week agentic build, and architectural production time fell 65 percent, against a target of 50 percent. Before any person noticed, a three-agent adversarial review had caught a wrong setback. Later we found that same model carried a footprint problem, traced to another project's plan set. The overlay rule and the blind pass are there for that reason.

Results

Held-out accuracy is the one accuracy claim Kanopi makes. Across three single-family bids priced blind, the mean error was 18.6 percent and the worst bid was 31.9 percent (June 10, 2026). We re-run the figure as the corpus grows.
An 11-bid blind backcast produced no miss that traced to takeoff quantities, and the building area of a 121-unit hotel landed within 1 percent.
A full quantity takeoff dropped from about 40 hours to about 40 minutes. Delivery takes 24 to 48 hours, blind pass and overlay included, and the internal cost is under $350.
On a 36,153 SF campus, the first proof from the vector rebuild caught a five-times overcount. The interior partitions came to 1,834 linear feet, against the 9,470 that an earlier desktop estimate had carried. Eleven contradictions between sheets went to the architect as requests for information, and an independent blind pass landed within 2.4 percent.
Eight builders and developers price their own work through Kanopi as of September 25, 2026, and over $600 million in proposals has gone out nationally in 2026.

The lesson

A number tested on its own training data is not an accuracy claim. Only a figure held out and forced to guess blind counts.

Public record

Kanopi

Stories from this work

A little better each dayA line from my high school basketball coach became the thesis behind every venture I have run.Catching my own five-times overcountTo prove the rebuilt estimating engine, I re-ran a project we had already priced the old way and found our own error.Nothing ships on one passEstimates and takeoffs are what I sell, so every number gets a blind second pass before it goes out.Selling the safe product, gating the risky oneI sold estimating immediately and held architecture back until my own building proved it.Sending an unverified bid to a competitor, and owning the mistakeI sent a bid to a competing contractor before reviewing it, and the comparison exposed my own error.The bug I built the gate forA markup error slipped into an early estimating-engine bid, and we used it to design the checks the engine still runs.The formula underpricing framing by six and a half timesAn audit of our estimating engine found a framing formula far below real cost, and we refused to patch it with a guess.The job posting that caught usWe applied to a posting that was really a lead funnel, then ran the same play for Common Ground.The markup stack that broke in MayA May 2026 restaurant bid shipped with a broken markup stack because I trusted the engine over my own review.The VP who asked, 'Can we buy this from you?'A materials manufacturer ran three blind takeoffs against Kanopi, and we caught our own errors before he did.

Questions about Kanopi, answered

Firm record

Common Ground keeps its own account of this engagement: Kanopi on the Common Ground wiki.

Back to all projects