By hand, it is slow
A submission set is dozens of sheets. Someone reads every room stamp, measures every wall and types hundreds of values into a spreadsheet. It takes days, and a slipped digit looks exactly like a correct one.
From submission plan to price, with every number traceable.
PlanSift reads Austrian building plans, from clean CAD PDFs and DXF exports to bad scans and phone photos, and returns values you can check: each one with its source box on the original page, its method, its calculation path and a confidence.
- Not found beats guessed
- Runs in your browser
- Demo uses synthetic plans only
The problem
The numbers are on the plan. Getting them out is the hard part.
With AI that guesses, it is dangerous
A model that fills every cell looks finished. But a plausible number without a source cannot be checked, and one invented area flows straight into the quantities and the price.
PlanSift: evidence or nothing
Every value points back to the place on the plan it came from. What cannot be read is reported as not found, with the reason, and is never filled in.
The product
Four steps from the drawing to the price
Each step leaves a trail you can follow back to the plan.
-
Read
Room stamps, opening tags, title blocks and room-book tables are parsed from the text layer where there is one, and read by vision models where there is not.
-
Measure
Room outlines, perimeters, gross floor area and facades are measured from vector geometry or DXF, at a scale that has been cross-checked.
-
Calculate
Derived values, quantities and bill-of-quantities positions come from formulas that name their exact inputs.
-
Explain
Every result opens into its sources, its calculation tree and its confidence. Constraints flag what cannot be true.
What you get, level by level
Factsheet
The official Excel factsheet: 100 parameters (78 mandatory, 22 bonus), each with its method and source.
Room book and quantities
A Raumbuch row per room, the openings, and areas and volumes such as BGF, NGF and BRI.
Bill of quantities
Positions with quantities and deduction rules, each traced to the rooms and openings behind it.
Price
A price range with every assumption labelled, and only where a sourced price basis exists.
In the viewer
Every result opens up
What you can do in the live demo, on synthetic projects, without installing anything.
“How was this calculated?”
Click any value: the calculation in words and in numbers, every input with its file and page, and “Show on the drawing” jumps to the exact box.
The 100-parameter checklist
All 100 official parameters, 78 mandatory and 22 bonus, with the answered count and filters. Correct a value and the original stays visible; dependent values and costs update, and one click restores it.
Document map
Every document of a project on one canvas, with the links between them. Tracing a value focuses the exact source box on its sheet.
A cost estimate you can take apart
A complete estimate in which every estimated quantity is marked, a cost tree by cost group with finish standards, scenarios such as another material or one part of the building, and an optimiser that keeps to the rules.
Excel with live formulas
Derived cells are real Excel formulas, so the workbook recalculates when an input changes. The filled workbook is checked against the official template.
Scans, photos, review queue
Scans and phone photos are read by a local vision model; what it cannot read stays not found. Recorded runs replay from the cache, and doubtful readings wait in a review queue, ranked by what depends on them.
3D model with self-checks
The rooms and walls read from the plans become a 3D building in the browser, and its geometry checks itself. More below
Accounts and run limits
The demo needs no account. Invited people sign in with an e-mail link to upload their own plans and start a run, within a daily run limit.
3D model
From the plans to a 3D building you can check
Rooms, walls and openings read from the plans become a 3D model right in the browser. Nothing is modelled by hand, and every part still points back to its source on the sheet.
Explode, walk, cut
Pull the storeys apart, walk through the building, or cut the walls open to look inside.
Colour by what matters
Colour the rooms by zone, unit, storey, review status or confidence, and see at a glance where the model is weak.
Self-checks that catch extraction errors
Overlapping rooms, broken outlines, storeys that do not stack and printed areas that disagree with the outline are flagged, with the values to blame. A check only warns; it never edits a value.
Export
IFC, glTF and DXF, for BIM tools, 3D viewers and CAD.
Demo
Watch a typo get caught, traced and fixed
A scripted walk-through of the real workflow on a synthetic project: read, trace, check, fix, export.
Read the demo as text
What makes it different
A number you cannot trace is a number you cannot use
Traceability, all the way down
Click any value and see the box on the original page, the calculation tree and the confidence. Review it, find the source of an error, fix it, and watch everything that depends on it recalculate.
- Value + unit25.52 m²
- SourceEG.pdf · page 1 · box · “25,52 m²”
- Methodread · measured · calculated · assumed
- Calculation pathNGF = Σ room_area
- Confidence99 %
Not found beats guessed
An empty cell is a finding; a wrong number is a liability. Checks may only lower a confidence. Nothing ever silently repairs a value.
Vector plans, DXF and BIM exports, scans, phone photos
A hybrid of text parsing, vector geometry and vision models. A weaker source never overrides a stronger one, and two sources that disagree become a visible conflict.
Constraints that catch the impossible
Ratio bands, area identities and geometry rules catch overlapping rooms, typos and OIB / ÖNORM violations, and point at the most likely cause.
Plan comparison and regulation check
Compare two versions of a plan room by room, and check a plan against formalised regulation items. Every finding says what it is based on.
Try it here
Every number knows where it came from
A made-up one-storey building from our synthetic demo set. Hover, tap or tab through the values.
Configurable inference
Your models, your machine
Vision models are asked only for what text and geometry could not answer. Which model that is, and where it runs, is your decision.
Local models
Qwen on your own hardware, through any OpenAI-compatible endpoint. Your plans never leave the building.
Claude
Anthropic’s Claude for the hardest scans. Off until you switch it on; every call is cached, logged and capped by a budget.
Your own endpoint
Point PlanSift at the endpoint your organisation already trusts. Remote hosts must be allowed explicitly and must use HTTPS.
No model at all is a valid choice: text and vector plans are read deterministically. The viewer runs in your browser, and a project you drop into it is processed locally and never uploaded. Plans reach the server only when a signed-in user uploads them for a run.
How a confidence is assigned
| Where the value comes from | Starting confidence |
|---|---|
| A correction you made in the viewer | 100 |
| Printed text that matches an unambiguous pattern | 95 |
| A measurement from vector geometry at a confirmed scale | 85 |
| A calculated value | weakest input × 0.95 |
| A vision model | below text and geometry; two independent passes must agree, otherwise the value is dropped |
| An assumption | at most 40, only from an explicit rule, and never used for quantities or prices |
Validators and rule checks can only lower these numbers; only independent sources that agree add a little. The confidence is a documented heuristic for deciding what to review first, not a probability.
Measured on the hackathon's private test data
- Mandatory parameters delivered
- Delivered values that are correct
- Confident-wrong values
- Values with a source box or a calculation path
Measured by the team on the private test projects of the hackathon; the plans themselves are confidential and are not shown anywhere on this site.
Technology
A Rust core that also runs in your browser
The same evaluation engine runs on the command line and, compiled to WebAssembly, inside the viewer. That is why a correction recalculates instantly, without a server.
- PlansPDF · DXF · scan · photo
- Extractorstext · geometry · vision
- Observationsvalue + source, nothing else
- Corereconcile · derive · check
- FactsheetExcel · viewer · BoQ · price
The pipeline, stage by stage
Scale, legend, page classes
Every sheet is classified as vector, scan, photo or mixed. The printed scale is cross-checked against dimension chains, and it belongs to one view, never to the whole page.
Text layer
A deterministic parser reads room stamps, opening tags such as F 120/120, title blocks and room-book tables. Precision over recall: ambiguous means nothing.
Geometry and DXF
Closed room polygons give area and perimeter at the validated scale, facades come from the elevations, and DXF entities are read exactly.
Vision, to fill gaps only
Asked only for what is still missing. A quality gate decides what may be read at all; two independent passes must agree exactly, and the second never sees the first answer.
Reconcile
Your correction beats text, text beats DXF, DXF beats vector measurement, and all of them beat vision. Equal sources that disagree become a conflict, and the value is withheld.
Derive, check, quantify
Formulas, ratio bands and exact rules run on the reconciled values. Incomplete sums are marked as partial and never feed a ratio, a quantity or a price.
- Rust33 crates, one contract between extraction and evaluation
- WebAssemblythe core and the rule engine run client-side
- No dependenciesthe viewer is plain HTML, CSS and JavaScript modules
- Deterministicsorted, seeded, no clocks in values