The platform
Software you don't install. A workforce you train.
Most AI tools are prompts wrapped around a model. Kanopi is built on generative skill learning: agents that acquire drafting and estimating ability the way an apprentice does, by working under a master, having every move corrected, and never making the same mistake twice. Our apprentice compounds across every project, forever.
Who trains it
Not trained on the internet. Trained under masters.
Every skill in the engine traces to a named human master working in their own discipline, on real projects, with real consequences. This is the moat: you cannot scrape what was never written down.
30 years
Real professional architects
The drafting skills are trained under licensed architects three decades into practice, captured in their own workflow, on permit sets that actually get sealed and submitted. Apprenticeship at machine scale, with the master's hands on the wheel.
$300M+ built
Master estimators
The pricing engine is calibrated by estimators who have bid, built, and reconciled hundreds of millions of dollars of completed construction. Every rate in the library traces to work that went vertical, not to a cost book's national average.
Always on
Live orchestrators
Every engagement runs under a human orchestrator directing the agent team in real time: sequencing the work, catching the edge cases, owning the client relationship. The machine compounds. The judgment stays human.
The Skill Engine
Six steps, one loop.
From a master's working session to a verified, compounding library of professional skill. This is the full loop, end to end.
Observe the master
A licensed architect drafts the way they always have. Underneath, Kanopi captures the work at the operation level: every Revit command, every sequence, every correction, streamed through a 138-tool programmatic control surface for the full Autodesk environment. Not screen recording. The actual action stream, with the geometry it produced.
BIM-native capture · full action telemetry
Encode into skills
Sessions are distilled into skills: versioned, executable procedures with preconditions, tolerances, and the code sections they satisfy. A skill is not a prompt. It is a tested unit of professional judgment, written down precisely enough that an agent can run it and a human can audit it.
Versioned like source code · auditable line by line
Verify against reality
Every skill runs a gauntlet before it enters the library: deterministic geometry checks, egress and clearance math, rated-assembly rules, IBC and IEBC gates. An agent must produce the same plan twice from the same inputs before we trust it once. Probabilistic output, deterministic acceptance.
Code-grounded gates · reproducibility required
Compound the library
Skills compose. Wall placement becomes unit layout. Unit layout becomes a floor stack. A floor stack becomes a full permit set with sheets, schedules, and annotations. Every new project regression-tests the whole library against everything it has ever drawn, so ability only moves one direction.
Composable hierarchy · regression-tested on real projects
Seal the loop
A licensed architect reviews and stamps every set. Every redline they make flows back as a training signal, corrected at the skill level, not the output level. The library gets the lesson permanently. This is the loop: the master teaches, the system compounds, the stamp stays human.
Human seal · every correction becomes curriculum
Run it on our own buildings first
We don't train on hypotheticals. The Skill Engine drafts our own projects, with our own capital at risk, before it touches anyone else's. On our own building in Midtown Phoenix the Design agents built the Revit model and caught and fixed their own errors with no human in the loop. That standard is the product.
Calibrated on owned projects · skin in the game
Why nobody else has this
Language models guess.
Buildings can't.
Getting a probabilistic model to act inside a deterministic discipline is the hardest problem in applied AI. An inch is not a token. A bearing wall is not a suggestion. This is what we spent the hard years building.
- The action-space problem
- A chat model picks the next word from a vocabulary. A drafting agent picks the next operation from an effectively infinite space of geometry, constraints, and code requirements, where one wrong move invalidates a hundred downstream decisions. We solved this with skills: pre-verified procedures that collapse the space to moves a master would actually make.
- The verification problem
- There is no autocomplete for life safety. Every output has to clear deterministic gates: dimensional tolerance, egress math, structural grid discipline, code citation. We built the gauntlet before we built the agents, which is the opposite order from everyone else, and the reason our errors get caught by software instead of by plan reviewers.
- The knowledge problem
- The internet knows what buildings look like. It does not know how an architect decides. That knowledge lives in working sessions, redlines, and thirty-year habits that have never been written down. Our capture pipeline is the only one we know of that harvests it at the operation level, with the practitioner's consent and participation, inside their own workflow.
- The calibration problem
- Design and price share one model, so every drawing decision must survive contact with a real budget. Our estimating engine is anchored to projects that were bid, built, and reconciled against firm cost. When the agents draw, they draw inside numbers that have already been proven on job sites, not inside a cost book's averages.
The result is a system that cannot be copied by training on the public internet, because the thing it knows doesn't exist on the public internet. It exists in the hands of the masters we train under, one verified skill at a time.
Inside Kanopi Estimate
Four things every other tool can't do.
Autonomous takeoff
Drop a plan PDF. Vision AI reads the sheets, classifies them, extracts quantities, automatically. You review, you don't redo.
vs. competitors: PlanSwift, Bluebeam: hours of manual click-to-measure
Calibrated to your shop
We learn from your past bids, your subs, your overhead, your history. The same project bid by you and bid by your competitor produces two different numbers, both correct.
vs. competitors: Sage, STACK, ProEst: generic RSMeans, identical to peers
Live market pricing
BLS Producer Price Index feed, daily 1build material prices, live tariff watch on imported categories. Quotes update with the market, not with the next quarterly software release.
vs. competitors: Most catalogs update quarterly, often months stale
Confidence on every line
Each line shows where the number came from: quoted by a sub, measured from the plan, assumed from history, or budgeted as an allowance. Owners trust transparent bids more than perfect-looking ones.
vs. competitors: Industry standard: opaque lump-sum totals