Personal AI NAS Command Deck
The service bench for a personal AI NAS
NestPulse sits beside your drives and your local models the way a service panel sits on a well made appliance. Four working planners live on it: capacity math, model fitting, library cleanup, and automation rules. Every number is computed in this tab, every dataset is a labeled demo, and nothing ever leaves the page.
- Storage Doctor Capacity, efficiency, and rebuild windows for five RAID levels
- Model Fit Quantization ladders against a RAM or VRAM budget
- Tidy Plans Duplicate hunting over a labeled demo library
- Rule Forge Conflict detection with a logged dry run
- Runs Offline
- Zero Dependencies
- Synthetic Data
- MIT Licensed
Storage Doctor
Pick your drives and rebuild speed. Every level shows usable space, efficiency, fault tolerance, and how long a rebuild would stretch.
A TB on the box is not a TiB on the disk, so usable space also shows in TiB. Rebuild estimates assume one spare writing a single stream at the chosen speed, with a 1.25x parity penalty on RAID 5 and 6. Real controllers differ.
Model Fit
Set a RAM or VRAM budget and a context length. Each model shows the largest quantization that fits, with the arithmetic spelled out.
| Model | Params | Best Quant | Weights | KV Cache | Total | Verdict |
|---|
Weights use each quantization's bits per weight. The KV cache assumes fp16 with two tensors per layer. Runtime overhead is a flat 1.00 GB. Specs are planning approximations from public model cards, not measured checkpoints.
Tidy Plans
Demo DatasetThe planner works on a labeled demo library with planted duplicates, burst sets, and aging screenshots. Lower the confidence threshold to see what was hiding.
Confidence bands: identical content hashes rank at 100 percent, burst sets at 75 percent, copy suffix patterns at 65 percent, and aged screenshots at 45 percent. The planner reads only this page's manifest, never a real disk.
Rule Forge
Demo DatasetNine demo rules run against three days of synthetic hourly sensor logs. The detector flags contradictions, shadowed rules, duplicate effects, and references the logs cannot satisfy, then checks each prediction against what actually fired.
Rules on the Books
Dry Run Results
Potential Conflicts
Validation Notes
Fires per Rule
| Rule | Fires | First Fire |
|---|
A contradiction means two rules can drive the same device in opposite directions at the same moment. Predicted is not observed: the results call out when a possible clash never actually happened in the demo logs.