Change the
approach.
The interesting problems often sit between disciplines. I look for a better representation, a constraint I can remove, or tools that become useful together, so that an awkward task becomes possible.
Ten changes of approach.
Some of these are shipped products. Others are experiments with open assumptions; their write-ups say which.
A checkable design program
Rather than accepting plausible geometry as a finished design, the compiler turns typed intent into CAD and independently rechecks the resulting artifacts. Transactional edits and derived assembly relationships keep model proposals inside explicit constraints. Simulation remains separate from proof that a physical part works.
An editable voxel world
Street panoramas supply appearance, map footprints supply volume, and a voxel grid supplies editability. Combining those representations creates a street that can be mined and rebuilt. A geometry diagnostic view exposes where the photographic illusion stops matching the underlying shape.
A shared company workspace
The product brings a company workspace together with local provider accounts and development environments. That creates a difficult boundary: shared coordination must not silently become shared credentials or permission to publish. Enrolled hosts, scoped connectors, and approval of the exact proposed change make that boundary explicit.
Requirements read once, reused everywhere
Job requirements can be extracted once and reused across candidates; a candidate’s facts can be reused across incompatible application forms. Separating those jobs makes matching easier to inspect and avoids paying for the same interpretation at every pairing. Submission receipts close the loop between a planned application and what actually happened.
A round guide and a slot
The alignment experiment compares two round guides with a round guide and a slot. The slot allows one direction of error while retaining yaw control. The method is to isolate that mechanical hypothesis in an unpowered fixture and measure it before adding motors or more complicated control software.
Predict locally, rewind when wrong
Client prediction and rollback reconciliation separate the feeling of responsiveness from authority over game outcomes. Shared simulation and separate physics, bot, and network tests help identify which layer is responsible when the game feels wrong.
Explicit outcome coverage
Related market titles do not establish equivalent settlement rules. This experiment models outcome coverage, fees, depth, and sequential execution together, then stops when execution state is uncertain. Its contribution is a way to inspect and test an apparent opportunity, not a claim of profitable trading.
Playback that adapts to the device
Browser capabilities, codec support, network behavior, and remote controls all affect the same viewing experience. Source selection, range caching, and FFmpeg fallbacks address those differences below the interface, while persistent progress keeps recovery from becoming a restart.
Scene ownership and a reviewed handoff
Scene ownership, durable coordination, and isolated drafts turn a collection of agents into a writing workflow. A frozen candidate is reviewed before selected files reach the source project. This makes continuity, recovery, and editorial control part of the system rather than relying on the agents to remember them.
An incentive that has to be tested
Listening requirements, timestamped prompts, and review credits address the tendency of a feedback exchange to reward shallow participation. The retrospective also records where execution fell short: successive technical foundations delayed a real musician trial. The lesson is to test the central social assumption earlier.
The portrait on the home page.
One generated image became a voxel bust you can turn. A second model drew the missing angles, Hunyuan3D reconstructed the shape, and a voxelizer rebuilt it from solid cubes. The shadows painted into the reference images were fitted and removed, so the browser lights it live.