Startup track
Built by Luciano Bogomolni
Construction projects often end with a walkthrough and a long list of defects to fix. Punchwalk turns that inspection into a documented punch list. An inspector films with Ray-Ban Meta glasses or a phone, explains what's wrong, and points to the issue. Each item includes a photo, the relevant moment in the video, the inspector's words, and a proposed fix with the responsible trade.
Muse Voice Transcribe creates timestamped transcripts, while SAM 3.1, supplied with clips through the Meta files API, follows pointing gestures and outlines defects. Muse Spark Open writes and verifies the entries against their photos. On a return visit, it compares new footage with the original evidence to review claimed fixes. Items without enough visual support wait for human review. The submitted demo processes uploaded footage; direct live capture through the Wearables Device Access Toolkit is a future step.
Built by Roman Fomin
Formwright helps people design small devices that do not already exist on a store shelf. A request for a leak alarm or plant sensor becomes a schematic, circuit board, enclosure, and firmware, with a route to production through Axo, the builder's existing 3D-printing business.
Muse Spark 1.3 selects components, creates the device design, inspects renders and writes firmware. Muse Spark Open acts as an independent critic, while SAM 3.1 supports measurements and visual checks. Muse Wonder assists with difficult designs, with Muse Image, Muse Voice Transcribe and Muse Spark 1.2 contributing additional capabilities. Code-based checks cover the board, enclosure, and firmware. In the demo, those checks catch incorrect components and wiring, then send the problems back to the agent to fix.
Built by Dodzi Agbenorku
When independently designed building or machine components occupy the same space, someone has to resolve the clash. ORCA goes beyond listing those conflicts. It proposes changes to the actual 3D model, tests them with geometry checks, and presents verified options for a person to approve.
Muse Spark through Meta Model API examines rendered views and model measurements, then proposes edits. Muse Spark Open handles bulk triage, while Muse Wonder takes a longer second attempt when the initial search finds no verified fix. Muse Image visualizes a checked design option, and Muse Voice Transcribe enables spoken requests. Muse Code can connect through ORCA's MCP server to run workflows using the same checks. The user reviews the before-and-after model, and the server checks the change again before applying it.
Built by Jimmy Yao and Andrew Tang
Voluma helps scientists turn micro-CT scans into traceable 3D measurements. Agents segment bones, identify regions and landmarks, perform measurements, and validate their work. Researchers can inspect the scan in slices or 3D and request corrections, with a record of the files and regions behind each result. It is a research tool, not a medical device. Muse Wonder through Meta Model API coordinates the work and runs specialist agents that write and execute Python. Those agents choose slices, boxes, and points for SAM 3.1, using its tracking across scan slices to label a volume. The team fine-tuned the open weights of Meta Segment Anything Model 3.1 to identify a growth-plate structure the base model did not detect in their scans. Muse Spark supported development tests and a comparison baseline. The project shows how agent workflows and specialized open weights can work together in a scientific task.
Built by Robin Kroonen
Glimmer Scope gives developers a way to inspect and steer concepts represented inside Muse Glimmer. Robin trained sparse autoencoders, tools that separate model activity into interpretable features, across nine layers of the open-weight model. The result is a browsable atlas and a demo where selected features become controls for changing how Muse Glimmer responds. Muse Spark 1.3 through Meta Model API labels features from examples of where they activate and evaluates whether steering changes the subject of a response while preserving readable language. Muse Glimmer itself is the subject of the analysis and runs the interactive demo. Turning up a character feature, such as a pirate captain, changes the response without adding a system prompt or fine-tuning the model. The project also exports steering directions for use with a quantized Muse Glimmer model.