§00prelude
this is a personal site, but it is also a probe. the explicit objective is to assemble a workflow that can replace me — first in skillset, eventually in solutions thinking. every project below is a step along that gradient.
i build at the seam between architectural technology, automation, and ai. i compose teams that are part human, part agent. the work that comes out the other side has to be real, deployable, and measurable, or it does not count.
§01the replace-me thesis
- every workflow i hold is a single point of failure.
- if a workflow can be described, it can be delegated.
- if it can be delegated, it should be — to humans or agents, whichever is cheaper at the required quality bar.
- the work that remains is what the system could not yet describe.
- that work is the only work worth doing.
§02method
record. transcribe. compress. test. ship. retire. the loop is boring on purpose. what matters is the rate at which a captured workflow stops needing me to run it.
record ──▶ transcribe ──▶ compress ──▶ test ──▶ ship
│
┌─────────────────────────┘
▼
retire (success)
observe (drift)
§03projects (active)
four shipping. one paused. each is documented to the point that the next maintainer — human or not — can pick it up cold.
#001archtech-os[running]
AEC workflow primitives, agent-orchestrated.
a runtime for architectural production tasks: mesh repair, parcel analysis, packet generation, fabrication output. each task is a typed function an agent can call. agents are scheduled by a planner that consumes natural-language project briefs.
rhinograsshopperpythonmodallanggraph
#002team-builder[running]
composition engine for hybrid human/agent teams.
given a project brief and a roster of humans + available agents, produces a team composition with explicit ownership, escalation paths, and a budget. re-composes as load shifts. used internally for every engagement.
brief ─┐
├─▶ planner ─▶ team(humans, agents) ─▶ commitments
roster─┘typescriptpostgresclaudegpt
#003john-clone-v3[training]
persistent operator agent · long-horizon delegation.
an always-on agent with read/write access to inbox, calendar, docs, and the codebases it maintains. handles routine correspondence, drafts proposals, and escalates anything outside its calibration. currently held at v3 while we tighten the held-for-review threshold.
claudecustom ragcronhuman-in-loop
#004workflow-mirror[running]
screen-record → reproducible playbook.
records a session of me doing a task, produces a step-by-step playbook with branch points, and stubs out the agent code that would execute it. the codification rate is the only metric that matters here.
ffmpegwhisperclaudeduckdb
#005spec-to-shop[paused]
natural-language → fabrication-ready geometry.
paused pending a better evaluation harness. the demos look great; the failure modes are silent and expensive. holding until we can guarantee tolerance on edge cases.
three.jsopencascadeclaude
§04how to use this site
- read the project entries — they are the actual work, not summaries of it.
- poke the live demos linked in each entry. they all run in-browser.
- ask the agent at /repl — it has read everything here and will answer in my voice. mistakes are mine, not its.
- subscribe to /rss for engineering notes; the cadence is irregular and unedited.