Michael Wahl

This very site is the reference implementation. As AI assistants increasingly mediate how people find and judge professionals, a site built only for humans gets misread. So this portfolio was engineered as a four-layer, agent-native stack from a single source of truth: an llms.txt discovery file and an expanded llms-full.txt knowledge document, a schema.org JSON-LD entity graph (Person, ProfilePage, Book, and every project cross-linked by @id), and an evidence layer of stable per-project Markdown URLs. It runs as a measured N=1 case study: an identical 25-prompt benchmark is scored across five AI surfaces (Google AI Mode, ChatGPT, Perplexity, Gemini, Claude) before and after the changes are crawled, on discovery, accuracy, completeness, citation authority, and whether the model can reconstruct the full evidence chain from person to production system to measurable result to source.

A proxy layer for AI coding agents that replaces naive full-repo context injection with a live structural model (import graph, symbol index, diff state) and builds a sub-8K-token context pack per request. It learns which files matter by parsing which symbols the model actually cites, biasing future context toward them.

A single local pre-flight suite for Claude Code — quality, dependencies, secrets, security review, and documentation gates behind one config file, producing one pass/fail report before a push. The deliberate design choice is to specify and freeze the config contract before building the runner: config files end up committed in other people's repos, so the JSON Schema (draft 2020-12) is validated and every gate prompt written against it first. Two pieces already execute — a tested bypass guard, a Claude Code PreToolUse hook that closes the `git push --no-verify` and agent-shells-out-to-git escape hatches with a loud, logged override, and a documentation gate proven end-to-end against live Confluence via the Rovo MCP (create → human edit → merge-update in place, human prose preserved). The remaining four gates are specified against the same shared contract but ship disabled until each earns trust on measured precision.

An enterprise manufacturer

An end-to-end pipeline for an enterprise manufacturer. Claude streams complete first-draft technical procedures from real engineering data (bills of materials, change notices), while a companion pipeline drives 3D tooling to auto-generate exploded-view illustrations and animated walkthroughs from raw CAD, including automatic revision-delta detection. Validated against real, large production assemblies with zero code changes between them.

An enterprise manufacturer

A Claude Code proofreader for large technical publications. It walks a PDF page by page applying brand and term rules, hyphenation, and callout-vs-label checks, using Opus vision to catch glyph-path and image-rendered defects the text layer misses, and auto-chunks PDFs over 32 MB. No API keys and no Python pipeline to maintain: tech authors run it from Claude Code and get a reviewer-ready findings table. A documented model comparison locks the tool to Opus after Sonnet returned zero findings where Opus caught real defects.

A persistent institutional-memory agent for engineering leadership. It stores decisions, incidents, and org knowledge and retrieves them with hybrid semantic plus full-text search (Reciprocal Rank Fusion). The thesis: the model is a commodity, and the accumulated institutional context is the defensible product.

An enterprise sales org

A production Claude plugin that puts prospecting, lead-qualification, and pipeline workflows directly in the assistant for an enterprise sales org. It's backed by a serverless broker that proxies CRM and data-enrichment APIs so credentials never ship inside the plugin, with per-rep authorization via an allowlist, full audit logging, and CI security gates.

A lightweight proxy that classifies each request and routes it to the cheapest capable Claude tier, with live cost tracking. A simple lookup goes to a small model; a hard reasoning task goes to a large one. A commercial enterprise edition adds policy controls, cost caps, analytics, and licensing.

Field service

A hands-free voice agent for service technicians that captures repair-order intake and warranty documentation by voice, then runs a real-time guard that flags the exact gaps that get warranty claims denied (vague language, missing codes, missing part numbers, missing diagnostic time) before submission.

A domain-agnostic monitoring agent that runs entirely on its own: on a schedule it captures an image, pulls in supporting context (weather, sensor readings, recent history), and has Claude assess the scene against a per-domain rubric using vision, then logs the assessment and alerts a human only when action is actually needed. The architecture is a reusable kernel plus swappable domain packs, so the same agent monitors a garden or a fleet of services with no code change, just a new rubric. It runs headless (no server, no daemon) on cheap edge hardware.

An opinionated open-source starter for running an autonomous AI agent under real security guardrails, built after onboarding an AI "employee" on a local machine. It codifies five principles as adaptable templates: the agent gets its own isolated service account (its own identity, not an extension of yours), credentials scoped to the one job it does, advise-don't-act with a human approving anything irreversible, a lean always-on agent with disposable scheduled sub-agents for heavy work, and a written agent handbook defining scope and escalation before day one. Ships sanitized config, handbook, and security-setup templates.

A local observability tool that reads the JSONL session transcripts coding agents like Claude Code already write to disk and turns them into a health dashboard. It computes six indicators: file-revisit rate (thrashing), turns-to-resolution, context/cache efficiency, whether the agent actually queries the project memory it built, a frustration index from user-side sentiment, and swarm detection separating orchestrator from sub-agent sessions. The point is to pinpoint where architecture is making the model thrash. It runs 100% locally: no cloud, no API keys, no telemetry.

An automated sales-intelligence agent that monitors a set of target accounts every day for two kinds of buying signal: hiring signals (scraping careers pages for roles that indicate demand for our services) and prospect intelligence (scanning business news via the Brave Search API across categories like product launches, electrification, regulatory events, and AI/ML adoption). For each signal it generates a personalized cold-outreach email draft with an LLM, tied to specific products, with false-positive filtering built into the prompts, then delivers a daily digest for human review before anything is sent. All signal-detection, prospecting, drafting, and state logic is custom, built on the open-source nanobot agent framework; it runs unattended on a daily schedule.

A team-enablement kit for engineering orgs adopting Claude Code as a coding partner, built on the premise that you brief the agent like a senior engineer joining the team. It provides a root CLAUDE.md briefing template (what and why, where things live, rules, flow, state), local per-directory CLAUDE.md files for sensitive modules like auth and infra, reusable skills for structured code review and systematic debugging, hooks for deterministic formatting, linting, and security scans, and progressive-context docs. It ships with a practical first-week onboarding guide covering setup, PR workflows, security scanning, and Jira/Confluence integration.

Prototyped native Android apps for Android 16 tablets using Google's Gemini models and AI Studio, exploring on-device and agentic app patterns where the OS itself can discover and invoke an app's capabilities from their documentation at runtime. A deliberate look across the frontier-model landscape beyond Claude.