RealityLabPROJECT BRIEF / META HACKATHONView code

BUILD FOR YOUR WORLD / ENGINEERING CURIOSITY

From “what is it?”
to “how does it work?”

For makers preparing hands-on demonstrations: photograph an everyday mechanism, explain one visible principle, then test a separate craft study.

Watch the guided exploration Try your own object

The guided exploration uses a labeled, recorded Meta run. Scanning your own object makes a new request after consent.

Actual recorded fan analysis presented as a conceptual blueprint with source-photo callouts
EVIDENCE WORKSPACE / CONCEPTUAL, NOT TO SCALE
THE PROBLEM

Seeing a part is easy.
Understanding its role takes more.

For a maker preparing a hands-on demonstration, a name alone leaves the useful questions unanswered: which part does what, what supports that explanation, and how could I explore the principle myself?

RealityLab keeps those questions together: photograph, evidence, component roles, visual study and a separate hands-on proposal. The aim is to reduce the jump between looking at an object and understanding it. Learning impact and demand still need user testing.

THE WORKING PRODUCT

One object.
A connected investigation.

Explore a source, challenge an inference, then follow the principle into a practical study.

  1. 01
    YOUR INPUT

    Bring a real object

    Take or upload up to three views. Photos are normalized and sent only after explicit consent.

  2. 02
    VISION + REASONING

    Reason with Meta

    Muse Spark Open identifies parts and principles, separating visible evidence from inferred mechanisms.

  3. 03
    LOCAL GRAPHICS

    Inspect the explanation

    Follow source-linked callouts, compare photo and edge studies, and export a conceptual blueprint.

  4. 04
    PREDICT / TRY / REPORT

    Test the explanation

    Meta proposes a separate craft study. Record what actually happens, then optionally send the exact report for possible explanations and distinguishing observations.

WHY META IS CENTRAL

The reasoning connects
the whole experience.

The first request reads the photographs. The second reasons over the validated parts and principles to create the construction study, or answer a scan-grounded question. An observation question can also carry the complete original step, prediction and reported result. Meta returns possible explanations and distinguishing checks, not a verified cause. The interface retains component references and report snapshots so users can inspect the basis of the response.

Inspect the server integration
ACTUAL INFERENCE CONFIGURATION

Muse Spark Open

meta-models/muse-spark-open-1026
Serving provider
Baseten
Request 01
Multimodal object analysis
Request 02, optional
Construction / learning / questions
Blueprint rendering
Local SVG + image-edge processing

Optional component and step illustrations use the Meta image API when configured. These are conceptual images, not repaired-outcome proof or photo-to-3D reconstruction. The interactive 3D fan is prepared geometry.

SHOW THE WORK

Real requests.
Inspectable results.

Five curated single-view examples, recorded on 28 October 2026. These demonstrate outputs; they are not an accuracy benchmark.

Times come from the saved server responses. They exclude upload and browser rendering and do not predict your next request. Model names, timestamps and request IDs are included in each JSON record.

Open the interactive examples

Inspect the evaluation protocol and known-failure case

WHAT WE LEARNED

Confidence needs
a visible boundary.

A valid JSON response does not establish a correct explanation. We kept the failures in the evaluation record and added a practical observation journal so users can compare predictions with actual results. These reports are not independent verification.

01 / RECOGNITION

A cropped tool was not identified correctly.

A severely cropped screwdriver produced uncertain tool hypotheses. More context or another angle is needed. This is a recognition failure, not evidence of universal support.

02 / PHYSICAL REASONING

The first craft proposal was mechanically inconsistent.

It expected a fixed straw to rotate and suggested unsupported airflow. The prompt was revised to prefer simpler studies and check mechanical consistency. The replacement remains untested physically.

03 / EVIDENCE

A photograph cannot verify hidden construction.

Internal mechanisms stay labeled inferred. Markers can be approximate; materials are hypotheses. Blueprints are conceptual, with no recovered dimensions or manufacturing tolerances.

Read the evaluation notes Open the physical-trial worksheet

BUILT DURING THE HACKATHON

Craft you can inspect.

The current Next.js implementation began on 28 October 2026. The current implementation is merged into main, which Render deploys after CI checks pass. Dependencies and photo sources are credited.

Validate before displaying

Strict result schemas, component-reference checks, bounded images, explicit consent and sanitized provider errors. Failed inference returns an error rather than a substituted sample.

Keep graphics grounded

Observed and inferred filters, editable approximate markers, source-linked callouts and controlled SVG exports. Photo-free downloads are the default.

Make the experience resilient

Cancel and retry, tab-memory lesson caching, reduced motion, keyboard controls, responsive layouts and WebGL fallback. Automated tests complement manual review.

START WITH SOMETHING FAMILIAR.

Leave with a better question.

Explore the recorded story Scan your own object

Physical build validation, broader object evaluation and user studies are the next proof points. No measured learning benefit or universal reconstruction capability is claimed.