Systems I've designed and built, most of them operated under Aidtake Labs. Each one is real, running infrastructure, documented as problem, constraints, architecture, and decisions.
Case studies
01CreativeProvenGeneration and review architecture is built.
02ApprovalHuman gateCreative approval, then a separate final approval. The approved artifact is hash-pinned.
03PublishProvenChannel identity checked, one-shot canary. One real public canary published.
04MetricsMeasuringCollected automatically for published videos.
05ExperimentMeasuringExperiment state is recorded against those metrics.
06DecisionWaiting for dataNeeds enough measurement windows. Nothing is decided automatically.
Auto-publication: disabledAny further publish needs a fresh human approval. Every publication is written to an append-only audit history.
Traffic OS
Private content-operations application for YouTube channel publishing
private · in production
Problem
Operate YouTube channel publishing without letting automation make judgment calls.
Constraints
Private, internal application — no third-party tenants.
Per-channel OAuth credential isolation; credentials never shared across channels.
Architecture
Live channel identity verification before any write action.
An immutable hash of each approved artifact, so what gets uploaded is exactly what was approved.
A human creative-approval step, then a separate, final publishing-approval step.
One-shot canary publishing controls — no batch or repeat publish without a fresh approval.
An append-only audit history of every publication.
Key decisions
No generic "force" or "bypass" path around any approval gate.
illustrative
AI agent operations system
Personal AI agent orchestration system and content pipeline
personal · in production, human-gated
Problem
Run autonomous agents against real infrastructure without losing accountability.
Architecture
Claude Code for the reasoning layer — reading state, deciding what to do, drafting the action.
n8n for workflow and integration glue between services and the agent layer.
A dedicated human-review UI for anything an agent does that gets flagged as drift.
A content pipeline, from planning to early measurement: Marcos (video editing, short-form and long-form), a daily blog pipeline with evidence, fact, narrative, privacy and site-QA gates, and human approval before anything publishes.
Key decisions
The blog stays on human approval: the strict autonomy acceptance test produced 2 of 3, and the gates were not loosened to reach 3.
Content lifecycle
Discovery and planningIn production
Recording and uploadIn productionValidated on one real job
Media (Marcos)In productionValidated on one real job
QAIn production
Publication gateHuman gated
Content and blogHuman gated
MeasurementMeasuringVery early search data
Revenue OSIn development Separate work in progress. Not presented as finished.
Where this stands
Last reviewed:
Current status
The blog pipeline runs daily. A person approves the blog and the final video; neither publishes alone.
Last meaningful milestone
October 2026: a real long-form job went from about 1.8 GB of raw video to a finished 1080p video of about 8 minutes, with audio mastering, motion graphics, YouTube metadata, chapters and three thumbnail candidates.
What's proven
A resumable large-video upload (about 1.8 GB) completed on one real video.
The video pipeline finished one real long-form production job.
Safe cleanup of local media after publication, validated end to end on disposable production jobs.
The daily blog pipeline discovers, writes, runs its checks and prepares articles.
What's still not proven
Publishing without a person. The acceptance rule was 3 of 3 and the result was 2 of 3, so production stays on human approval.
A track record: the long-form video result is one job, not a series.
Whether any of this brings readers or viewers. Search data is still very early.
Why the third autonomy run stayed on human review is not a general result: it is one run.
Next meaningful validation
Observe real daily production runs and real drafts under human approval. If auto-publishing is reopened later, it would use the same strict 3-of-3 acceptance rule.
No independent cross-brand reference for paint sheen and finish equivalence existed.
Constraints
Independent — no brand affiliation implied.
illustrative
CompatSense
Connectivity and compatibility intelligence tooling
in active use · limited public detail
The only public detail available today: connectivity and compatibility intelligence tooling. There isn't enough published material yet for a full case study, so this stays a short entry rather than a padded-out one.