I Don't Advise on AI. I Build It.

Framework

The 39-agent autonomous software factory: 8 departments, a 10-stage looping pipeline, and git-based governance that shipped mobile apps end-to-end. The framework itself was archived in June 2026 after an adversarial review and a written retirement postmortem; the apps it shipped stay live, and its governance patterns moved into a machine-wide action-boundary hook.

System

A C#/.NET 8 port of my phone agent's call loop, built to test the design: G.711 audio, local voice activity detection and barge-in, tool calling, retrieval, caller memory, verbatim recorded disclosures, and call-frequency rules for payment reminders, with 137 tests. I ran the same scripted call through four voice models (xAI, OpenAI Realtime, OpenAI GPT-Live, Gemini Live); xAI replied fastest. Testing it against live models found three bugs in my production line, fixed the same day.

App

Zero-question personality profiling from real-world behavioral signal, no quizzes, no self-reporting. Live on the Apple App Store with resumable full-history import and credit-based pass-through pricing.

App

iOS nature-soundscape app built on bioacoustic synthesis models, synthesizing sleep/wellness audio from scientific data, not loops. Live on the Apple App Store.

Product

Autonomous bookkeeping on a commercial-grade ledger: bank sync, OFX/QFX/CSV history import, universal receipt ingest with auto-split, and an auditor that resolves uncertain categorizations under a stated tax posture instead of parking them for a human. Flagged or auth-failed bank connections are retried three times before any email goes out. Escalates to a council when it is genuinely unsure. Multi-tenant, 211 tests.

Product

A decision aid for someone choosing about hospice on another person's behalf. Built from the published research rather than from reassurance, with a sample brief rendered through the same path that produces a real one, so the output can be read before any details are entered.

System

The cold-email engine for Agor AI, built and deployed in one day. Sender v2 runs fast, follow-up and cold lanes behind a compliance gate with RFC 8058 one-click unsubscribe. ReplyAlert detects replies and honors opt-outs, auto-replies never count as replies, and a follow-up sequencer sits on top. A HALT switch and a laptop heartbeat dead-man switch stop it when I am unreachable, and a new sending domain warms up on a code-enforced ramp.

Tooling

Turns a blog post into an 8-second editorial motion graphic in the live site's brand, replacing the static cover on every third post and carrying the same motion into the social posts. A director picks from a catalog of archetypes, and a QA gate checked by a mutant suite and a backtest rejects any frame whose numbers do not match the post. Live on scored.tools and modelstack.digital.

Framework

A producer that opens pull requests and a review council that decides whether they land, pointed at the whole portfolio rather than only its own repo. Quorum is tiered: code an agent wrote needs unanimity, code a human wrote needs a majority. The merge path was proven end to end on a canary before it was trusted, four swallowed failures were found and fixed, and the privileged merge token stays out of the proposing agent's reach.

Framework

A pilot of the action-boundary governance idea built in eight phases on a sample billing module: a permission floor, a Stop-time claim check on a commit-stable fingerprint, test protection and scope checks, a premortem gate with a tree-aware retry budget, subagent handback checks, and per-rule enforcement scored by a 60-run eval.

Agent

Gated inbox-triage agent that drafts and, when armed, sends on my behalf. Ships OFF, then DRAFT, then LIVE. A structural never-act class fails closed, an away-safe posture changes its behavior when I am unreachable, and a phone command inbox lets me steer it from a text message.

Agent

A daily job-application pipeline that found, by measurement, it had filed nothing in 125 roles, then was rebuilt to kit only what it can file. It now files Greenhouse forms end to end, reads the emailed security code, answers form questions by reasoning from standing facts, and routes Ashby to an attended lane because Ashby flags unattended submits as spam.

Agent

Vertical outbound pipeline for a medical-coding client: keyless sourcing from public provider registries, a compliant human-reviewed outreach queue, and a CAN-SPAM-guarded email channel hardened by three adversarial reviewers before it was armed. 92 tests. Put on ice on 24 September 2026 with its tasks disabled.

Framework

A 16-stage pipeline that plans and drafts a five-book series end to end: series bible, per-chapter contracts, deterministic checks instead of model self-grading, and resume-from-checkpoint after a usage-limit stall stranded 110 chapters. Currently paused at 40 of 150.

Book

An 824-page book of 25 AI-to-AI Socratic dialogues (~313k words), composed by orchestrating persona-specific claude -p subprocesses. $0 incremental compute.

Book

A 50-chapter, 25,951-word book taken from spine and ledger to an audited draft, a published reading edition, and a print-ready 5.5 x 8.5 interior.

Writing

Three works taken in a fixed order, premise artifacts, then novella, then treatment, then screenplay, so nothing reaches the script that the prose has not already earned. THE SLOW ANSWER is a 77-page feature at 11,633 spoken words; GOODWILL and GOOD COMPANY followed the same derivation, and one was rewritten in a second voice to see what survived the change.

Writing

A complete archive of all 105 Bill Gross Investment Outlooks (1978–2026), paired with a Claude skill that drafts new outlooks in his voice and reasoning.

Research

A pre-registered study of LLM agents trading a market they are also moving: hypotheses frozen before any run, three agent arms in a sterile harness, cross-family judges, and a red-team pass that forced the headline comparison down to preliminary and underpowered. Published with a Zenodo DOI; the arXiv submission is staged pending endorsement.

Patent

A provisional patent application filed pro se (64/087,364), with a parametric CAD pipeline that renders draftsman-grade hidden-line figures straight from the design spec, and a complete nonprovisional package prepared behind it.

Film

A 23.5-minute explainer film on special and general relativity, every scene rendered from code, with a council-reviewed script that fixed six physics errors before a frame was made, synthesized narration, and a captioned master.

Audio

A multi-voice radio drama taken from script to finished master by pipeline: a structured script model, casting by measured vocal pitch rather than by ear, synthesized performances, an original score, built foley, and a ducked mixdown that has to clear a ship gate before release. The pipeline is reusable for audio plays, narrated essays, and multi-character audiobooks.

Audio

Guided training audio where the timing is the product, not a byproduct of how long the narration ran. One timeline source drives the voice render, a tempo-locked music grid, and region-based bed mixing, and a master ships only after it clears three gates, corrects its own true peak by measurement, and proves every cue lands on the beat.

Tooling

A repeatable pipeline for dialogue-driven commercials: characters whose mouths say the actual words, scene-to-scene continuity, ducked score, and native 16:9 plus 9:16 masters. Produced the Agor Agents launch film and two ModelStack spots that ran as a live Meta campaign.

System

A personal knowledge brain on Postgres + pgvector, 28,000+ pages, a live MCP server, and automated collectors for email, calendar, notes, and bookmarks. Nightly offsite backup and a chunk sweep that catches pages filed but never indexed, because a page nothing can retrieve is not knowledge. Full-text search and the link graph were both repaired in August 2026 after one root cause left each returning less than the brain actually held.

Tooling

A multi-model router that sends self-contained subtasks to the cheapest model that can do them, with per-model cost attribution in the status line, cache-aware routing, and a master off switch.

Tooling

A local coding agent on opencode plus Ollama that runs with no API spend, benchmarked on the machine it actually runs on before being trusted with real work rather than adopted on impressions.

Tooling

A local read-only dashboard for what every AI subscription actually costs and consumes, plus a System health tile answering the question a cron fleet cannot answer for itself: what would tell me something broke?

Tooling

An encrypted bundle of my Claude Code configuration with a one-line bootstrap, so a cloud session carries the same tools, skills and APIs as the session on my own machine.

Tooling

An autonomous iOS release pipeline: archive → upload → poll App Store Connect → TestFlight, with a self-healing known-failure-mode autofix catalog.

Plugin

A published Claude Code plugin: /orient, /ideate, /spec, /scaffold, /build, /ship-web, /ship-mobile, /launch, /learn, plus /start and /start-next for resuming the arc from saved handoff state, orchestrating the full build-ship-learn loop across the portfolio from one command.

Plugin

A Claude Code plugin that assigns code a production tier, ranks audit findings by expected failure cost, and records deliberate deferrals with the trigger that should promote them. Backed by a 74-note corpus covering 111 topics and a detector suite that was corrected by running it against a real repo.

Skill

A Claude Code plugin that reads the live conversation, recommends the next chain of skills, subagents, and slash-commands, then executes it immediately, with inline confirms kept only on high-risk steps.

Template

A scaffold for Stripe-billed Claude Code skills, the reusable pattern for packaging and selling agent skills with metered access.

Skills

100+ personal Claude Code skills and slash-commands that encode how I work, email voice, an anamnesis memory protocol, self-learning, adversarial councils that span model families, audio-drama and video production, prescriptive timed audio, bulk marketplace listing, document production, and cross-repo orchestration.

Site

Stealth product, parked deliberately at the June 2026 portfolio review. SEO and infrastructure wired and dormant ahead of a launch decision.