The model proposes, a person commits, and every call is logged
A safety gate so the AI can suggest actions but never takes a risky one on its own.
Key decisions
- Fail-closed: deny unless explicitly permitted
- Checkpoint / resume with human approval
Selected Work
Difficult systems, the engineering decisions behind them, and the constraints they had to respect.
Each piece is labeled for what it is.
- Professional Experience
- Systems built in professional roles.
- Architecture Pattern
- An approach described on its own terms, independent of any one deployment.
- Internal Build
- Blue Pixel’s own work.
A source-traceable account read a reviewer can trust in 30 seconds
An AI account summary where every number traces back to the source, or it says "no data."
Key decisions
- Every figure traces to the system of record
- Contradictions surfaced, not smoothed over
AI-guided intake that scopes a small project without the consultant bill
Guided questions, not a blank text box, with a human in control.
Key decisions
- Typed interactions, not free-text
- Human-gated submission
Commitments made on a call become structured records, each human-confirmed
Call transcripts, scrubbed of personal data, turned into clean records a person confirms with one click.
Key decisions
- PII redacted before the model sees it
- Typed events, verbatim-grounded, human-confirmed
Right to be forgotten, on top of a log that's never deleted
Destroy one key and a person's data becomes permanently unreadable, with the history left intact.
Key decisions
- Per-person key destruction = permanent erasure
- Build-time guard blocks PII leaks
One app with clean seams, split only when the evidence says to
Skip the cost and complexity of microservices until there's a real reason to pay it.
Key decisions
- One deployable now, splittable later
- Contract-shaped seams at external boundaries
Have a problem this shape?
Start with the problem. The right engagement follows from there.
Start a Conversation