Arthur Locke
AI Automation Engineer
AI Automation Engineer with roughly two years of hands-on workflow automation experience — from n8n pipelines to integration workflows — and, over the last several months, hands-on work building multi-agent LLM systems and MCP servers. Treats prompt engineering as a production discipline: versioned, context-budgeted, and tested against real outputs before anything goes live. Has shipped automations across Claude, GPT, and Gemini and knows when to reach for which. At Keen Software House, led company-wide LLM adoption, built MCP servers giving non-technical staff natural-language control of live automation infrastructure, and architected a three-agent pipeline that generates and self-heals test scripts autonomously.
Projects
- Built an enterprise multi-agent AI platform across the full Microsoft stack, evaluating a Microsoft-native service (Copilot Studio, Power Platform, Azure OpenAI) against custom pro-code for every architectural capability before defaulting to custom code — 15 documented ADRs recording the reasoning, not just a stated principle.
- Multi-agent Copilot Studio design: a parent agent using the Connected Agents pattern routes each conversation to one of two specialized sub-agents — a Knowledge agent (RAG over an indexed corpus) and an Enterprise Integration agent (Teams/Outlook/SharePoint tooling) — verified live through real conversations, including catching and documenting a genuine platform-side OAuth consent bug in the Outlook connector across three separate investigation attempts.
- The one deliberate pro-code layer: a Semantic Kernel API behind Azure API Management, Entra ID-authenticated with role-based authorization enforced end-to-end, calling Azure OpenAI and a dedicated Azure AI Content Safety moderation layer via managed identity — zero API keys anywhere in the platform.
AzureBicepCopilot StudioPower PlatformSemantic KernelEntra IDGitHub Actions
- Built a self-hosted AI support-ticket triage and classification platform with Claude treated as a genuinely replaceable intelligence provider, not a hardwired dependency: strict separation between Chatwoot (unmodified), two self-hosted MCP servers, and the LLM layer, documented across 6 ADRs recording the actual reasoning.
- Three interchangeable ways to run the identical classification logic: an automated FastAPI webhook pipeline calling the Claude API directly, an on-demand CLI reusing the same workflow code with no persistent server, and a fully API-key-free mode where a Claude Code session (or a scheduled routine) drives the MCP servers' tools directly — validated by manually classifying and triaging real seeded tickets through MCP before the automated pipeline was even built.
- A permanent, compliance-driven architectural boundary rather than a v1 limitation: no code path anywhere in the system ever sends an AI-generated message to a customer directly, regardless of confidence score, real or self-reported — captured as its own ADR immediately after confirming a demo "known issue match" had been simulated, not computed, and following that honesty through to its actual design consequence.
PythonFastAPIChatwootClaude APIPostgreSQL (pgvector)Docker ComposeGoogle Cloud Run
- Built a multi-package platform where Claude Code itself is the agent driving two custom MCP servers (TestRail, Unreal Engine 5.8) to migrate manual regression suites into AI-evaluated, executable tests — deliberately no server-side LLM API call in this pipeline; the interactive evaluate/generate steps are guided by versioned markdown instruction templates.
- Deliberate scoping decision, not the default "wire an agent to both MCP servers and let it improvise": the LLM's judgment is confined to the one place it genuinely can't be hardcoded — deciding whether a manual case is automatable given what's actually discovered live.
- AI evaluation step scores each manual test case's automatability (confidence 0-1) against live-discovered capabilities — run for real against 9 seeded cases: 7 correctly automated, 2 correctly rejected for the same proven engine-input gap. Human-approval gate is a real enforced function call, not a documentation reminder.
PythonPydanticuv workspacemcp SDKSQLiteDocker
- Built a self-hosted observability platform for AI coding agents themselves — ingests real, native OpenTelemetry telemetry from Claude Code, Codex CLI, and Gemini CLI (each vendor's actual documented schema, deliberately no synthetic data), normalized into one shared per-agent-adapter ingestion pipeline.
- OpenTelemetry Collector receives OTLP over gRPC/HTTP into a Next.js 15 dashboard (sessions, cost, token usage, tool calls, per-session timeline); verified end-to-end with a real, live-connected Claude Code session — not a test payload.
- Optional AWS path: 5 Terraform modules mirroring telemetry into CloudWatch via EMF metrics — deployed and destroyed for real against a live AWS account across two regions, verified each time with live AWS CLI calls rather than trusting Terraform's own state.
Next.js 15TypeScriptSQLiteDocker ComposeOpenTelemetry CollectorTerraformGitHub ActionsAWS CloudWatch/IAM
- Built a 9-package TypeScript monorepo: contract runner CLI, structural drift detection engine, AI drift analyzer, and 4-tool MCP server.
- AI layer: claude-sonnet-4-6 with prompt caching — static spec pinned in the cache, dynamic drift report injected per run — produces semantic severity analysis and remediation actions. Prompt caching chosen deliberately to minimise token cost on repeated runs.
- AI test generation (--ai-test-gen): Claude generates edge-case TestCase objects from the internal spec at runtime and merges with spec-driven tests. Full CI: 24/24 unit tests and 8/8 contract tests passing.
TypeScriptFastifyZodAJVVitestDockerGitHub Actions
TestRail MCP — MCP Server for TestRail Test Management
- Built a from-scratch MCP server wrapping TestRail's REST API v2, mirroring the UseResponse MCP's dual-transport architecture (stdio + Google Cloud Run streamable-http) and OAuth 2.0/PKCE + signed JWT security model.
- Two-tier tool-gating: five category flags (read, case/run/result writes, deletes) stacked with an optional named-tool allowlist that can only narrow — never widen — what a category flag permits.
- Always-on list_tool_policy introspection tool; automated test suite validates registered tools match policy across representative flag combinations.
PythonFastMCPhttpxDockeruvicorn/Starlette
- 5-phase agentic pipeline: Planner (converts PRD to structured test plan) → Generator (produces Playwright TypeScript specs) → Debug (resolves real failures) → Healer (simulates selector drift, auto-repairs broken locators).
- 39 tests across 5 suites, all passing; real bugs diagnosed and fixed during the debug phase. 10 CLI playbooks for repeatable, CI-ready agent execution.
PlaywrightTypeScriptClaude Code
- Developed a Next.js 15 single-page template for professional and business use (TypeScript, Tailwind CSS v4, shadcn/ui, Credly certification badge integration), with environment-based builds configured for Meta Business Verification readiness — the platform this multi-facet site itself is built on.
Next.js 15TypeScriptTailwind CSS v4shadcn/ui
Skills
LLMs & Prompt Engineering
Versioned, context-budgeted promptsprompt caching (Anthropic)structured outputstool useRAGmodel-tier selection (Haiku/Sonnet/Opus, GPT/Gemini)
Agentic Systems
Multi-agent pipeline design (Planner/Generator/Healer)MCP server developmentOAuth2/JWT-secured streamable-HTTPruntime capability discoveryhuman-in-the-loop designfallback/escalation pathsClaude CodeLM Studio
Workflow Automation
n8n (self-hosted, queue mode)multi-step workflowsconditionals & error handlingApifywebhook pipelines
API & Integration
REST APIsOAuthJSON parsing & schema validationpaginationrate-limit handlingOpenAPI contract testing
Scripting & Dev
PythonTypeScriptNode.jsBash
Governance & Security
Prompt injection awarenessleast-privilege agent designsecret handlingaudit trailsISC2 CC
Cloud & Infra
OCI (Compute, OKE, IAM, Gen AI Services)AWS (EC2, S3, Lambda, CloudWatch)KubernetesDockerTerraformGitHub ActionsOpenTelemetry
Enablement
Cross-functional AI adoptionhands-on mentoringinternal documentation & playbooks
Experience
AI, Automation & Development — Freelance
2023 — Present- Article Generation Pipeline: multi-stage Python pipeline — RSS polling → deduplication against a processed-items database → context search for prompt enrichment → Claude API with prompt caching (static context cached, dynamic content injected per run) → Airtable human-in-the-loop review → published markdown output.
- RAG Chatbot: retrieval-augmented generation chatbot (OpenAI, Gemini, Supabase vector DB) with multi-channel delivery via WhatsApp and Telegram webhooks — full lifecycle from integration design and prompt engineering to client deployment.
- Tour City Interactive Chatbot: stateful n8n conversation logic with PostgreSQL-backed session state for a tour company's city walking quest; QR code generation for frictionless entry, message routing and branching across separate user journeys.
- Accounting Workflow Automation: end-to-end tax return automation — WordPress form → n8n webhook → Python processing → Czech government-regulated XML and PDF with schema validation for regulatory compliance.
Advanced Quality Assurance Specialist — Keen Software House a.s.
Aug 2020 — Jul 2026- Led company-wide AI adoption: evaluated tools, selected Claude over existing ChatGPT usage, defined the rollout plan for GWS + Claude integration, authored onboarding documentation, and delivered hands-on mentoring to engineers and non-technical staff.
- Connected the internal n8n instance to Claude via MCP server — team members can observe, diagnose, and create automation workflows through plain English with no technical knowledge required.
- Built and deployed the UseResponse MCP server — dual transport (stdio locally, streamable-http on Google Cloud Run for scheduled Cloud Routines); OAuth 2.0/PKCE with signed HS256 JWTs; seven independently-gated tool categories, deny-by-default.
- Architected the SE Tester AI Agent System — a three-agent pipeline (Planner interprets spec files and game source; Generator produces executable XML test scripts; Healer auto-diagnoses and repairs failures from MCP-provided logs).
- Designed and shipped n8n automation workflows for spam filtering and automated weekly report generation — saving ~5 man-hours every week with zero ongoing maintenance cost.
Certifications
Oracle Cloud Infrastructure 2025 Certified Generative AI ProfessionalOracle Cloud Infrastructure 2025 Certified Architect AssociateCertified in Cybersecurity (CC) — ISC2API Test Automation