QwertyS · HTML Challenge 2026 — first place · AI Builders Hackathon — participated, no prize
Open the hackathon app · Watch the submitted demo · Read the Devpost submission · Public hackathon source · Professional pilot results
▶ Watch AEDEXA — Devpost AI Builders hackathon pitch and demo
The original recording is hosted on Nurkhan Aimukatov's YouTube channel. Click the preview to watch the hackathon edition. The animation below is the landing-page identity; it is separate from the submitted product demo.
The landing page introduces AEDEXA through a rotating 3D portal tetrahedron. Its faces open onto architectural scenes, connecting the geometric mark to the product's movement between drawings, site geometry and spatial models. The wordmark, topographic background and alternating drawing-led sections carry that same visual language into the rest of the page.
The original 26-second landing animation, converted to GIF for GitHub. Open the retained WebM animation.
See the full hackathon landing page
Captured from the public hackathon host on 12 October 2026. The statistics are the project's published checks, not a newly performed independent benchmark.
A capture of the hackathon edition. All product images and public demonstrations in this repository belong to that edition, including its earlier interface iterations.
Early architectural decisions bring several kinds of information together: a parcel, a topographic survey, utility lines, surrounding buildings and a proposed structure. Those facts are often scattered between CAD files, maps and individual drawings. AEDEXA puts them into a browser workspace where a user can move from a drawing or map to a site model, inspect the assumptions and export the result.
The central idea is that a useful model must remain explainable. A measured elevation, a regulatory value and an AI-inferred detail have different meanings. Missing information should stay visible rather than becoming an apparently precise number.
| Workspace | Input and result |
|---|---|
| Placement and setbacks | A DWG/DXF site drawing, calibrated plan image or map outline leads to a boundary, constraints, buildable area and building placement. |
| DWG topographic survey | Survey geometry, spot elevations and contours become classified objects and a triangulated 3D terrain surface, with DAE/DXF export. |
| 2D to 3D reconstruction | Source views feed an audited parametric recipe, deterministic geometry construction and a separate model review. Model access is required. |
| Project archive | Saved results can be inspected and reopened rather than disappearing after one calculation. |
A satellite view can establish where the site is, but a traced polygon does not by itself prove cadastral boundaries or surveying accuracy. The application carries the working boundary into its calculation and inspection views. CAD layer roles and drawing type help determine whether an uploaded sheet actually represents a site suitable for placement.
Earlier hackathon interface: the parcel contour remains a visible, editable input.
An overhead plan is useful for understanding the survey structure; a 3D surface makes changes in elevation easier to inspect. AEDEXA exposes both views and the geometry used to build them. DWG decoding runs through LibreDWG WebAssembly, while terrain construction uses triangulation of the available elevation evidence.
Topographic processing in the dark hackathon interface.
Compare the earlier 2D and 3D terrain views
These captures document the earlier hackathon interface; they are not screenshots of a separate commercial edition.
An isolated model on an infinite grid hides the context that matters in site planning. The placement view joins the proposed building to the working parcel, terrain and constraints. Setbacks and utility protection areas contribute to the usable area; surroundings and building dimensions remain part of the inspection.
A functional hackathon capture of the placement stage.
The browser reconstruction workflow audits the drawing's source views, creates a structured recipe, checks it against a schema, constructs geometry and runs an independent visual review. The point is to make unsupported geometry distinguishable from measured CAD information. AI checks are explicit actions rather than a prerequisite for every terrain or placement operation.
Hackathon edition: the darker reconstruction workspace.
Inspect earlier building and overhead examples
Multiple viewpoints make footprint and height relationships easier to question.
The web edition uses React 19, TypeScript, Vite/vinext, Three.js, Leaflet, LibreDWG WebAssembly, Delaunator and polygon-clipping. A Cloudflare Worker serves the application; Cloudflare D1 supports accounts and project synchronization. Geometry and terrain calculations run in the browser. Optional AI analysis and reconstruction use server routes and configured model-provider access.
flowchart LR
A[Drawing or map input] --> B[Drawing type and layer roles]
B --> C[Boundary, elevations and constraints]
C --> D[Buildable area and terrain]
D --> E[Building placement and inspection]
E --> F[Saved project and export]
A --> G[Optional reconstruction audit]
G --> H[Recipe, solver and model review]
H --> E
The registry attaches source documents and clauses to planning rules. Results use
statuses including PASS, FAIL, MISSING_DATA and EXPERT_REVIEW. This makes incomplete
or unverified inputs visible. It does not turn a hackathon application into an approved
construction-design or regulatory certification system.
The Devpost submission records 275 automated checks, a regression corpus of 69 third-party DWG drawings (63 readable models) and 19 source-linked building-code rules. These are the development results reported for the submitted web edition, rather than a new test execution performed for this showcase.
AEDEXA received first place at HTML Challenge 2026. The organizer's official results, published 19 September, name Qwertys in first place. It was also submitted to the AI Builders Hackathon on Devpost; that submission did not receive a prize. These are separate events and outcomes. The hackathon record links the event information, submitted project and video, official result and retained team certificates.
Official HTML Challenge results
Original organizer artwork from the 19 September results publication. Credit: HTML Challenge / @html_challenge. Preserved without alteration.
The implementation includes local setup, geometry tests, demonstration drawings and deployment instructions. The hosted app and submitted video refer to the hackathon edition. This showcase distributes presentation materials; implementation access is managed separately.
What the earlier Python drawing audit established
The earlier Python investigation audited 66 industrial DWG files. One produced a built model; other cases stopped at conversion, view discovery, view pairing or supported geometry checks. This was a diagnostic experiment on that older pipeline, separate from the browser platform shown above. It is retained as development history rather than used as a performance claim for the current web workflow.
| Earlier audit outcome | Files |
|---|---|
| Conversion failure | 2 |
| No usable view, despite captions | 21 |
| No usable view and no captions | 23 |
| Plan/elevation pairing failure | 2 |
| Unsupported or invalid archetype | 17 |
| Model built | 1 |
AEDEXA completed a closed pilot with Ansar Architects, one company, from 20 September to 4 October 2026. Four professionals worked with five sites: three current projects and two archived cases. The report covers timing, drawing compatibility, geometric comparisons, rule checks and usability.
The four timed workflows were 56.7–88.0% faster than the reported manual baselines, including the stated review and corrections. Reliability remained a separate concern: 8/11 DWG files opened fully, 47/50 definite verdicts agreed with the lead project architect, and one false PASS was recorded. SUS was 66/100 (n=4). Read the complete pilot results.
Trial use continued from 5 October; results for that stage are not reported yet. The expanded implementation remains private. The public product gallery and video show the hackathon edition; the pilot chart visualizes reported aggregate measurements and is not a customer-workspace screenshot.
1st place at HTML Challenge 2026, organized by Allinone. The certificate names Akhmer Shakhnazar and team Qwertys. This award belongs to HTML Challenge; the separate AI Builders Hackathon submission did not receive a prize.
Original documents and certificate gallery.
View the certificate
HTML Challenge 2026 — 1st place, Qwertys
Shakhnazar Akhmer / Eye172 and Nurkhan Aimukatov / pip00sya.
Companion showcase on Nurkhan's account · Shakhnazar's projects · Nurkhan's projects.
The two showcases document the same joint project. Product facts, authors and evidence are shared; each account maintains its own presentation repository.