The photo work created a different kind of technical problem: too much visual material and not enough structure. Years of images, professional fight photography, personal references, screenshots, DNGs, and phone photos only become useful when they can be searched, scored, and reviewed.
Where it is now
The local person finder GUI selects reference-person folders, scans roots and drives, runs local face recognition with multiple CPU workers, logs progress, checkpoints work, and outputs matches, low-confidence non-matches, errors, and JSONL result records.
On top of that, the photo tagger uses CLIP-style local tagging, category assignment, and relative 1-to-5 scoring. For the professional CV photo set, the scoring was recalibrated so only about 2 percent of images become 5-star candidates.
The barriers
The barriers were scale and ambiguity. Face matching has false positives. Image tagging can over-label. Professional usefulness is category-specific: a social photo, business portrait, fight image, and creative image should not be judged by the same absolute scale.
How I did it
I used a staged approach: identify candidate images, tag them locally, apply relative category scoring, and then use the strongest assets where they fit. The personal site now uses real images of me for credibility and generated system visuals for abstract infrastructure sections.