Systems & AI Engineer
I’m Louijie, a systems and AI engineer with 14+ years of experience. Here you can explore my public code, working examples and the decisions behind them.
Evidence operations desk
A dated source review, not a live dashboard. Open each file to inspect the retained output and the question it leaves open.
AeroLoop / PR #21
500 ms recovers in the tested profiles. Longer gaps expose the boundary.
accepted: falseHeld-feedback baseline, seeds 0/1/2. Mixed means some mission passes; all one-second paired checks fail. This is not a safe-duration limit.
Open limit. Later prediction and landing-guard studies still retain accepted: false.
Inspect source ↗Omniweft / PR #22 / PR-015
Workers can fail without gaining mutation authority. Explicit retry can recover the original receipt.
| Phase | Status | World revision |
|---|---|---|
| timeout | timed_out | 0 |
| cancelled | cancelled | 0 |
| crashed | crashed | 0 |
| superseded | superseded | 0 |
Recorded statuses from the public archive. Prevented proposals leave the world at revision zero.
Open limit. The parent and volatile policy host must survive. No OS sandbox claim.
Inspect source ↗LudoWeave Engine / M244 / PR #262
Pause, step, seek and resume with visible tick and playback status through M244.
Open limit. Presentation controls do not change recorded input. Seeking has no latency guarantee.
Inspect source ↗Selected public work
Follow a recorded flight, inspect a saved room or explore what an AI tool sees. Each project links the output to its code, decisions and current limits.
Control experiments you can inspect
A simulation-only drone laboratory measuring sensor-outage recovery and capture-aware landing. The retained stress studies report accepted: false, with replayable flight evidence and explicit failure boundaries.
A world engine controlled with Python
An early world engine with bounded worker lifecycle and explicit retry recovery. Late proposals cannot mutate the world; uncertain delivery can recover the original receipt without duplicating an edit.
3D scenes for AI tools
A 3D engine that lets software and AI tools change a scene, then inspect images and object data from the same scene version.
A real example from Cogniform
A person can look at a picture and recognize an object. An AI tool also needs information it can work with: distance, surface direction and object identity.
Cogniform produces these four views of the same scene, so a tool can connect what it sees to the object it can change.
See the example and how it was checked
Appearance
Distance
Surface direction
Object identityA note from Louijie
Louijie Compo
I work across the boundaries between physical devices, software platforms and applied AI. I make interfaces explicit, test assumptions early and carry ideas from architecture into working implementation.
These projects give you a way into that work: look at a frame, follow a request, or inspect what happens when a change is rejected. Pick the part that makes you curious.
Engineering journal
AeroLoop's retained simulation study recovers from 500 ms missing captures under the tested conditions. One- and two-second gaps expose a failure boundary that prediction and a landing guard do not close.
Read articleOmniweft keeps mutation authority in the parent and policy host. A crashed or superseded worker cannot revive a proposal, and explicit retry preserves the exact request when delivery is uncertain.
Read articleFollow writing and engineering notes
Add this feed to your RSS reader to follow new articles and engineering notes. No signup or tracking on this site.
Follow via RSSProfessional connection
For conversations about systems, architecture, applied AI, or carrying technical direction into implementation.