Inspect the interfaces and engineering decisions behind the work. These are captures of running development builds—not generated product mockups.
Video Factory · Actual local workspace, seeded with a fictional example. The capture shows planning UI, not a completed AI-generated production or a customer result.Inside the production workflow ↗
FirstCut · Working-title iOS development build. Real simulator captures; color bars are a synthetic export-test fixture. Not an App Store release.From edit decisions to exported video ↗
The transcript editor. Selecting text marks the video range; the footage is untouched until export.
How I built it
Deleting a sentence from the transcript marks the matching time range instead of touching the footage, so every cut stays reversible until export. Export composites in a single pass.
Why it could matter to you
Turn editing decisions into text operations while preserving the original footage.
Nine simulator tests pass, including a synthetic video exported with audio and captions and checked at five seconds
Cuts are reversible because the source media is never modified
Scope & limitations
Working title. Not on the App Store; real-device transcription, long-clip performance, and paid AI calls remain to validate.
Local development tool
YouTube Video Factory
A local workspace that keeps a video's script, takes, edits and render in one place.
Built for
My own YouTube channel, where the same project was spread across four unrelated tools.
One project view: script, transcript and edit recipe, with rendering driven from the recipe.
How I built it
Edits are stored as a recipe against the transcript rather than as a rendered file, so a change re-renders from the original takes instead of from the last export.
Why it could matter to you
Keep a repeatable production process attached to each channel and project.
Runs locally against a SQLite project database
Rendering is driven by FFmpeg from the stored recipe
Scope & limitations
Not a hosted public demo. No audience growth, publishing reliability, or time-savings claim has been measured here.
macOS development build
NexusTerm
A Mac terminal app that shows you what an AI coding agent is actually doing.
Built for
Myself, running coding agents all day and unable to tell live work from yesterday's transcript.
A screen recording of the activity view rebuilding after a partial transcript write, plus the XCTest run output.
NexusTerm has a written engineering account below; these descriptions are not a public demonstration.
How I built it
Activity is rebuilt from the agent's own transcript files rather than tailed live, so a half-written record is never displayed as a finished action. Records are only ingested once complete.
Why it could matter to you
Make agent work inspectable while separating historical activity from current work.
Native macOS app, built in Xcode from a Swift package
Regression tests cover every byte boundary of a Unicode fixture, file replacement, and stale activity
Scope & limitations
Not a release or live dashboard recording. Session matching, large-file performance, and Swift 6 package compilation still need work.
Independent engineering project
Sentinel
A permission layer that sits between an AI coding assistant and your files.
Built for
Testing whether an assistant's boundaries can be enforced in code instead of asked for in a prompt.
A terminal capture of a blocked path traversal and a redacted secret, alongside the audit chain it produced.
Sentinel has a written engineering account below; these descriptions are not a public demonstration.
How I built it
Path validation, approval checks and secret redaction run outside the model, so a prompt cannot talk its way past them. Every action writes a hash-linked audit entry.
Why it could matter to you
Make permissions and review explicit instead of relying on the model to behave.
Checks are enforced in application code, not in the system prompt
A redacted run log showing a verification failure escalating, and the architecture diagram from the case study.
ai-os has a written engineering account below; these descriptions are not a public demonstration.
How I built it
Each task carries a budget, a verification command that must pass before the task counts as done, and a heartbeat. A missed heartbeat escalates to a human rather than retrying quietly.
Why it could matter to you
Make unattended work inspectable and give failures an owner.
Private operating history is not an independently audited uptime or client ROI claim. Written up from a running system rather than a public repository.
App Store portfolio
Native apps on the App Store
Independent iPhone apps you can download and judge for yourself.
Built for
Shipping real products end to end, including the unglamorous parts: review, privacy, pricing and updates.
