THE STRUCTURAL LAYER FOR MODERN AI DATA SYSTEMS
Capture, validate, and ship robot-manipulation data with full operational assurance.
We work at the seam where physical data collection meets the training pipeline: the field rigs, the operations layer around them, and the memory infrastructure that keeps operators and agents working from the same record.
Capture. Multi-camera field rigs. Bluetooth pairing, network-attached camera control, synchronized start and stop across a cohort, tagging mid-take, and an ingest loop built so footage is never silently lost.
Validate. Claim-gated collection pools. Reserve, lock, auto-return, reconcile at end of day. Coverage is measured against target, not against optimism.
Ship. Verification is a first-class state. A capture is confirmed, unverified, or failed — and we never record the first without evidence.
| Repository | What it does | Stack |
|---|---|---|
| cortex | A private markdown knowledge base served to every Claude surface over MCP, with a read path that proves its own citations. | TypeScript MCP AGPL-3.0 |
| cold-read | Rewrites context-heavy text so a reader with zero background can follow it, and flags what is missing instead of inventing it. | Shell MIT |
| taskflow | Claim-gated collection pool, in production at tisktask.space. Internal tool; source private. | Next.js Apps Script |
| grain-community | The crowd's read on any video — community verdict on AI-generated media. Docs, releases, and feedback. | PWA Media provenance |
- Honest state over optimistic state.
unverifiedis a valid answer.confirmedrequires evidence. - Design for the after. Setup is 20 percent of a build. Day-2 operations are the other 80 — what drifts, what gets fat-fingered, what has to heal itself without a maintainer on the line.
- Fail loudly into a log. Silent failure is the only unrecoverable kind.
- Structure over decoration. Hairlines, hierarchy, and a single accent. Nothing on screen that is not load-bearing.
DATA. INFRASTRUCTURE. ASSURED.