AI Projects › Client playbook
If a client handed me their data estate tomorrow, whether a marketing department’s media library, a manufacturer’s engineering shares or an agency’s twenty years of drives, this is the engagement. Every step has already been run at scale, on worse hardware.
Six steps
1. Week 1: rules of engagement and a metadata map
Agree the risk ladder and approval format before touching anything. Mount every source read-only. Walk metadata only: names, sizes, dates. Deliverable: a sized estate map and a schedule set by the slowest device.
2. Weeks 1 to 3: fingerprint everything, in parallel
One hash pass per disk, all disks at once, written to a database in real time. Each volume passes gates: counts match, every file is hashed, a random independent re-hash agrees. Measured rates were 30 to 100+ MiB/s per spinning disk. Deliverable: a live console the client can watch.
3. Week 3: the copy map and the risk register
Copies per content, single-copy exposure by owner and project, reclaim potential in TB and dollars, and broken or unplayable files. This build found 1,517 videos missing their index atom. Deliverable: a ranked work order, not a slide. See the copy map.
4. Week 4: protect, then reclaim
Copy single-copy data to a second home first, verified by hash. Pick golden records by health and completeness. Reclaim only where a hash survives on at least two other disks, with written approval for each batch. Deliverable: a trim manifest and a ledger of every removal.
5. Weeks 4 to 6: compress with lineage and gates
GPU encoding with triage rules, quality gates and a metadata parity gate. Every compressed child is linked to its parent hash. Deliverable: space returned, with proof of quality.
6. Then: make it searchable
Hybrid retrieval over the clean library: path, transcript and visual similarity, with every result keyed by fingerprint. Only after that, evaluate whether fine-tuning adds value. Deliverable: a library a person or an agent can search without losing track of identity.
Weekly scorecard
- Percent of bytes fingerprinted.
- Single-copy terabytes, trending down.
- Verified reclaim, in terabytes and dollars.
- Defects caught before damage.
- Rows changed without approval (target: zero).
- Owner hours spent. The walk-away ratio so far is 30 minutes of decisions for 4.5 hours of safe work.
What I need from you
- Read-only access to the sources, and a named person who can approve.
- A rough list of what matters most to the business.
- Patience for week 1, because I will ask a lot of questions before I touch anything.
What is not claimed
Retrieval is designed, not in production. Fine-tuning is deferred. The agent message bus is designed, not built. The GPU lane is on hold. Client names are not published.