I build owned context systems for operators who are tired of explaining which business, client, or decision applies before useful work can begin.
The record stays in files they control. Software prepares the relevant evidence. The operator keeps judgment over sends, publishing, money, and production changes.
My commercial work is at Scale With Search.
You know the decision exists, but you cannot recover why it was made. A handoff stalls until you retell the history. An AI answer misses the one constraint that belongs to this account.
The repositories below show how I keep those contexts separate, searchable, inspectable, and useful.
| Repository | What the files demonstrate |
|---|---|
| subtlebodhi | A reference architecture for owned Markdown memory, bounded context loading, durable state, and reusable procedures. |
| pretool-memory | A Claude Code hook that retrieves relevant local knowledge before read-oriented tool calls. |
| session-ledger | A zero-dependency SQLite and FTS5 archive for local session records. |
| vault-crawl | A Rust retrieval core with provenance, content-addressed storage, search, and read-only diagnostics. |
| safe-api | A Python wrapper that places dry runs, scope controls, duplicate checks, circuit breaking, and receipts around REST writes. |
| observer-daemon | An external validator that checks model output against a written behavioral standard and records corrections. |
| observer-protocol | The written control model for approval gates, correction history, drift detection, and voice intake. |
| web2md | A single-command utility that converts a web page into clean, source-attributed Markdown. |
- Keep source material and correction history in systems the operator controls.
- Retrieve the smallest relevant context instead of loading an entire archive.
- Put human gates where software can publish, send, delete, spend, or change production state.
- Report what the system inspected, changed, skipped, or could not verify.
- Document each public boundary so the repository does not imply private data or deployed behavior that it does not contain.
The shared doctrine is published as seventeen principles of knowledge work.
Start with a $1,500 credited working session; scope and terms on the work page.
This profile README was written from the current repository files and runtime-backed portfolio records. Model assistance helped with implementation. The repository-specific README remains the authority for each tool's boundaries and provenance.