AI Projects › Specific projects
Everybody has had a time when their data grew out of control. Here are the specific projects, from personal to professional, that taught me how to fix it.
Clip sets for the people who matter
A birthday, a recital, a day on the water. The footage is spread across phones, cameras and drones, and the best moments are minutes inside hours. The work is finding the right clips (by time and GPS as well as filename), cutting them cleanly and delivering something people watch more than once. The method matters here too: originals are never touched, and every cut keeps a pointer to its source.
Video data redundancy
Footage had been copied forward across about twenty generations of drives. Fingerprinting 14.7 million files across 126 TiB showed how lopsided it was. See the copy map.
- 12.9 TiB existed exactly once, on single disks.
- 82.6 TiB of the total was theoretically reclaimable.
- The 777,579 contents stored six or more times took 36.5 TiB to hold 4.4 TiB of unique material.
Lesson: protect first, reclaim second, and prove every copy by hash.
Archiving, mining and sorting memories
Once the library was proven, the next job was structure: copying it into a final layout. A 2.9 TB copy ran overnight, driven by a 495-row sheet. A media probe across 683,000 files recorded codec, resolution, GPS and camera make, which is the groundwork for searching by what is in the footage. Compressed copies stay linked to their originals by hash.
Lesson: fingerprint first, then index. Retrieval fails when identity is based on file paths.
Inventory and understanding of past client structure
Old project folders hold the history of who you served, what you made and how you organized it. I rebuild that picture from the files themselves: sites, dates, cameras and deliverables, separating camera-assigned names (DCIM, 100MEDIA) from human-assigned project names. When 40 of 48 files matched, I asked where the other 8 belonged instead of accepting the number. They came from other shoots on a reused card.
Lesson: a match is a question, not an answer.
When “done” was not done
A rebuilt five-drive NAS showed that migrations marked complete were 6% to 24% complete. One share held 17.6 times more on the drive than in its migrated copy, another 12.4 times more, and a 168 GB share had only 398 MB migrated. A forgotten device holding 13 TB turned up, and a 2% sample showed 24,043 of 24,068 files already safe elsewhere.
Lesson: verify the result, not the exit code.
Protecting the work while it runs
Compression passed its quality gate and still dropped the 360 degree flag on 10 files. I found the cause within minutes, repaired all 10 losslessly and put a guard in every compressor script. The full account is on the two nights page.
← Types of projects | The journey →