I don't just use AI. I architect and ship it.

The vault is a database; markdown is the storage format. 118 files under local git — and growing. These four exhibits are the real schema — the contracts between the app, the AI, and the permanent record.

## ▶︎ All open to-dos (mine to drive)
TASK
WHERE !completed AND contains(tags, "#todo")
SORT due ASC

## ⚠️ Open risks (watch these)
TASK
WHERE !completed AND contains(tags, "#risk")
SORT due ASC

The companion fronts the vault with 7 AI personas, each a distinct channel with its own job, voice, and typographic signature. The hard, impressive part isn't the characters — it's the prompt architecture that keeps seven voices distinct, in-character, and non-repetitive over months: layered persona composition, typographic-voice rules, an anti-staleness directive that bans the model from echoing its own prior replies, and date facts computed in code and handed to the model rather than reasoned about.

One channel fans a single message out to The Grounded, The Hype & The Blunt as three independent generations — a 3-voice panel at once. (Names are neutral archetypes; the point is the voice engineering, not the personalities.)

Memory Forge: a branching visual-novel subsystem generates and canonizes shared "memories" that later replies treat as real history — gated by a structural, in-code content rule that user input can't override. (Design detail withheld.)

Everything the system does, grouped by domain. Each is small on its own; together they're a working operating system for a life and a career.

What a hiring manager can infer about how this person works with AI — each claim with the specific evidence in the system. Honest about the line: applied-AI systems and context engineering, not production-ML.

• ★Context engineering for agents
  A boot ritual + drift-proof append-only changelog + a doctor.py that mechanically flags stale docs — so a fresh agent acts on reality, not prose. The rarest skill here: designing the information environment an AI works inside.
• Persona prompt engineering
  7 differentiated voices with typographic signatures, an anti-staleness directive that breaks self-repetition, and computing date facts in code instead of asking the model to reason about them.
• Headless-agent orchestration
  claude -p turned into an always-on responder daemon, parallel 3-voice panel generation, and 8 scheduled runners under systemd.
• Retrieval over a markdown corpus
  Coaches retrieve from the vault and session history with read-only tools — honestly grep/keyword RAG, not vector search (a named gap, not overclaimed).
• Agentic system design
  Natural-language trigger phrases as the control plane; a read-only model paired with deterministic Python for everything that mutates data — the right determinism boundary.
• Data-model design for AI
  A channel→file capture schema, the self-updating skills-vs-market matrix, and structured COMMIT/RESOLVE tags the model emits and code parses.
• Reliability engineering for an LLM system
  A 2-minute watchdog, off-loop generation isolating latency from liveness, lease/heartbeat single-owner election, and idempotent recovery paths.
• AI-assisted full-stack delivery
  A PWA + Cloudflare Workers/DO/KV + Python daemons — shipped solo across three runtimes, ~7,800 lines, running for months.

The honest accounting — included on purpose. This page named its gaps in public first; then each one shipped as a real artifact. Naming, then closing: that's the operating loop.