The software I've built and shipped

Three portable .md artifacts, firewall-clean

WORKING_IDENTITY, IDENTITY_PRIMER, DEPLOYMENT_KIT — canonical portable context, deployed to the vault, consumed by the site assistant.

Next.js 16 module system, JSON file-database with live SSE, agent-assignment queue

Private 13-module operations console this site exports from: file-per-record JSON with live SSE, agent-assignment queue with a human-gated 'done'. Public template available.

Scheduled headless Claude Code + systemd timers + tier-scoped MCP

Autonomous nightly curation agent over the vault: routes material to sensitivity tiers, shelves atomic notes, weaves the link graph, maintains Maps of Content.

Hardened VPS + Cloudflare Tunnel/Access + MCP filesystem + git-crypt + gocryptfs

The serving layer of the shared brain: outbound-only tunnel, edge identity gating, three-tier encryption, GitHub sync across all surfaces.

Python engine + FastAPI service + Vite/React consoles

Genuinely-predictive order-confirmation analytics: forecasts which materials will miss confirmation next week, splits the at-risk book into recoverable vs structural, and hands the operator the specific lever to pull. Backtested across a run of historical weekly snapshots and validated on a held-out week.

JSON snapshots + nightly box-sync ingestion + Command Center console module

Usage tracking across every Claude surface: per-project token buckets locally, nightly ingestion from the remote boxes, daily snapshots, and a console module that renders spend and usage trends.

Always-on VPS, systemd-run agents, issue-driven dispatch, acceptance harness

An always-on agents box running autonomous Claude agents: a coding agent that answers task issues with pull requests, a Telegram personal assistant, and a shared runtime with its own verify harness — work continues while the operator sleeps.

Bash HUD wired into Claude Code settings, 30s refresh, gist-synced on every edit

The live session instrument: one line rendering model, context pressure, cost, and usage gauges inside every Claude Code session — with a toggle and a quiet mode, and a hook that publishes the script to a public gist the moment it changes.

Python engine over ERP BOM-explosion exports, versioned with anonymized test fixtures

Audits MRP master-data parameters across finished goods and components: ABC×XYZ segmentation drives geography-tiered policy targets (just-in-time through overseas), and every parameter that deviates from policy is flagged for analyst review with the dollar exposure attached. The structured migration of a proven console prototype into a version-controlled, distributable local engine — raw ERP exports never leave the machine.

Versioned markdown doctrine in a private git repo + session-start mirror-sync hook

The operating doctrine as infrastructure: one canonical, versioned NEXUS.md (any edit bumps the version) that a session-start hook pulls and mirrors into every Claude Code session — fail-loud, dropping a visible stale marker whenever the sync chain breaks. How the AI operates is governed from a single source of truth instead of drifting per surface.

Git-versioned ~/.claude with scheduled auto-sync commits to a private repo

The entire Claude Code home directory as a maintained, restorable build: session hooks, permission policy, the skill-library mount, the status-line instrument, sync scripts, and written failover + migration runbooks — auto-committed on a schedule so the machine's AI setup survives any single computer. The substrate the Claude skeleton map renders.

Single-file React artifact, in-browser XLSX ingestion, ported classification engine

Root-causes every unconfirmed order line in a weekly snapshot: the ERP's nested-IF classification waterfall ported to JavaScript and validated line-for-line against the reference classification, then ranked by cause category — supply, logistics, master data — with drill-down to the individual order. The proof that the classification logic could live outside the spreadsheet; its engine fed the confirmation tools that followed.

Single-file React artifact, embedded snapshot + XLSX ingestion

The descriptive ancestor of the forecaster: week-over-week order-confirmation tracking with drill-downs across the usual planning dimensions — served from an embedded snapshot or a fresh workbook drop, no backend anywhere. Its interaction patterns graduated into the predictive build; this is where they were proven.

Tauri v2, Rust, React 19, TypeScript — Windows + Android

Cross-platform PDF suite: viewer, page organizer, annotation, forms, OCR. One Tauri codebase shipping to Windows desktop and side-loaded Android.

MV3 Chrome/Firefox extension, vanilla-JS content scripts + vendored tokenizer, all data local

Live meters injected into claude.ai: per-conversation token count (exact when the stream exposes usage, a labeled approximation otherwise), a cache countdown timer, cache hit rate, and unrounded session/weekly usage bars. A modified build of the MIT claude-counter extension — this fork adds adaptive usage capture from the SSE stream, model-aware context limits, image/document token estimates, and a fail-visible badge when the page layout shifts.

Python 3.12+ (stdlib only, zero runtime deps), pytest ×48, ruff, GitHub Actions CI

Public BOM/MRP data toolkit: multi-level BOM explosion with quantity roll-up, where-used lookup, and production-plan churn diff over plain CSV. Ships with fictional demo data.

Node 22 + TypeScript + Ink, multi-provider (Anthropic, OpenAI, Google, xAI, OpenRouter, Ollama), MCP client, vitest + biome

A personal multi-provider CLI coding agent — Claude Code's shape, any model: frontier APIs or free local models behind one terminal UI, with tool use, MCP servers, and a real test suite. In daily use.

Cloudflare Worker + Anthropic API (streaming, cached) + React chat island

Site-embedded assistant in Ian's locked voice on 3 surfaces; rate-limited, spend-capped, 8/8 adversarial angles held including the employer firewall.

Astro + vCard 3.0 + logo-QR generator

Share-only landscape business card: save-contact vCard, tap rows, QR with the iQ mark baked in; noindex for privacy.

Python probability engine + FastAPI + MCP server + React operator console, one Docker container

Supply-chain order-confirmation risk service, built end to end: risk as a conditional probability table measured from week-to-week transition history, served over an API and an MCP server, with a React operator console on top. Fictional company, synthetic data.

Node + PWA, rebuilt live from disk on every build

A drillable reference map of the whole Claude Code meta-infrastructure — doctrine, config, hooks, memory, the skills library, plugins and MCP, companion repos, and every dev port. Reads live off disk, so the map cannot go stale on its own.

Next.js 16 + React 19 + Tailwind v4 + Anthropic API

Review-management dashboard for a small business: pulls reviews into one queue, drafts on-voice replies with AI, and tracks response health. Being rebuilt as a public demo on the AppliedIQ site.

Astro 6 + React 19 + Tailwind v4, Cloudflare Workers, streaming 'Ask the tooling' assistant (Anthropic API, SSE), Resend

17-page consulting/portfolio site: live supply-chain demos, services ladder, engagement-hub contact form, full 'Operational Instrument' brand system.

React island, multi-level BOM-explosion engine (per-unit quantities, where-used), synthetic data

Field-qualified BOM search, reverse where-used, multi-level explosion with true per-unit quantities across every usage path.

React island, week-over-week churn diff-engine (noise vs. real reschedules), synthetic data

Week-over-week production-plan volatility tracker: separates real reschedules from MRP renumbering noise, attributes changes to planner vs system.

React island, first-match-wins root-cause waterfall engine, synthetic data

Roots every unconfirmed unit to a cause via first-match-wins waterfall, drills category→cause→line, exports a ranked worklist.

React island, phantom-demand scoring + action-routing engine, synthetic data

Finds past-due production orders faking dependent demand, phantom-scores them, and sorts each into one action with where-used BOM impact.

React island, firmest-date arbitration + four-way triage engine, synthetic data

Daily inbound worklist: picks the firmest arrival date per open shipment and sorts each into verify/chase/reschedule/drop.

React island, per-bin integrity rules engine (offsetting-error detection), synthetic data

Per-bin physical-inventory integrity checks: impossible quantities, offsetting errors that net believable, line-side overstaging.

React island + Recharts, five-parameter grading engine (tier × ABC × XYZ, dollar exposure), synthetic data

MRP parameter-governance console: grades five parameters against sourcing-tier × ABC × XYZ targets, dollar exposure on every finding.

Claude skill: SAP CO PDF + ERP master-data extract + BOM extract + production plan → state-aware BOM validation and a dated implementation plan, two xlsx workbooks

Computes feasible implementation windows per Change Order with binding constraints, stock projection, multi-CO conflict report, readiness deadlines.

SKILL.md packages on claude.ai + Claude Code

The full house system: cognitive registers, operational skills, deploy machinery, and versioned skill/tool conventions — one library serving every Claude surface.