Anonymized portfolio piece. All system names, internal identifiers, and source content are removed. The mechanism, the pipeline shape, and the measured numbers are real.
The problem
Turning real project work into shareable case studies has a confidentiality problem baked in: the work that best proves capability is usually also the work covered by an employer or client agreement, or is mixed in with material that should never leave the source system at all. The naive fix, writing carefully and then proofreading, does not scale and does not hold up under scrutiny, because the person writing the draft is the same person checking it, with the same blind spots.
What I built
A four-stage content pipeline that mines a source repository for provable accomplishments and turns them into anonymized case studies, with a required adversarial gate between drafting and shipping:
- Mine. A fan-out of small, cheap agents each scan one slice of the source material in parallel and extract candidate accomplishments as short structured claims.
- Rank and synthesize. A single stronger agent dedupes and clusters the candidates, ranks them by strength of proof, and returns a short list. A human reviews the list and picks which items become case studies.
- Draft and fix. Each chosen item runs through its own small pipeline: one agent drafts the case study, then a separate agent is instructed to argue the draft is NOT clean, actively hunting for anything specific enough to identify the source: names, dollar figures, internal system names, anything personal. On doubt, it defaults to flagging a violation rather than clearing it. A fixer agent resolves every flag it raises.
- Repeat per source, per confidentiality bar. The same shape runs again against a second, differently-sensitive source, with the redaction agent instructed against a different bar for that material: describe the engineering, never the content the engineering handles.
The load-bearing design choice is that the adversarial check is a required pipeline stage; the draft cannot advance until the check passes. The agent doing the checking has no stake in the draft looking good; its only job is to find reasons it is not safe to publish.
The result
Across five drafts pulled from real source material, the adversarial gate caught and fixed four leaks that would otherwise have shipped: specifics that would have identified the source despite the draft otherwise reading clean. A human read every final file before anything was called shippable. This case study, and the file it sits next to in the same portfolio, are themselves output of this pipeline: proof by demonstration.
Why it transfers
The honest caveat matters as much as the result: this gate is an adversarial review plus a human sign-off. It is a strong filter, and a filter can miss. That distinction is the actual lesson: any pipeline that produces content from sensitive source material needs a review stage with an incentive to find problems, and it needs a human as the last check regardless of how good the automated pass gets. The pattern generalizes past writing: any workflow that mines a sensitive source and produces a shareable artifact benefits from separating "drafts it" from "tries to break it" into two different agents with two different jobs.
Josiah Bujanda. AI & Construction Technology Consultant. Mechanical engineer with more than ten years in commercial lighting controls and Division 26 electrical, plus hands-on architecture of production AI and automation systems.