AI Director and Applied AI Systems Architect. I build the AI systems a company runs on, and the layer that keeps them alive.

Principal, MLCx Consultancy · Chief AI Advisor, Fourth Ground Consulting

Sole architect and builder on every system described here. System design, data flow, automation topology, prompt engineering, integrations, deployment, and monitoring.

Before AI, nineteen years in digital marketing and growth operations, starting in 2007. I founded and scaled an outsourcing firm in Davao City from zero to over 100 full-time staff, then led the operational architecture behind a company's acquisition and its post-buyout integration. That is why my AI work starts with the process logic and not with the model.

Retrieval-grounded messaging assistants. A suggested-reply system for a multi-trade home-services group operating across several US metro markets: per-department voice modeling from the team's own historical conversations, a pre-generation and caching layer that removed a ten to twelve second response wait, and a nightly learning sync that keeps each department's voice current. Shipped as a Chrome Manifest V3 extension against a live n8n backend. A routing and caching layer cut cost per draft by roughly 43% on real traffic with no quality loss on substantive messages, and diagnosing a replay storm took a further 65 to 70% off daily running cost. High-risk messages are escalated to a human channel. Nothing auto-sends.

Call-time dispatch. A dispatch engine covering multiple regional markets: technician matching by skill and roster, calendar-aware slot offering built from live road distances, an operator panel used by customer service reps during live calls, and a passive learning layer that observes real rep behavior without ever replaying it into production. The routing engine is versioned and deploys without downtime.

Voice modeling at scale. A team-wide training console that recovered the per-department split a messaging API does not expose, embedded years of conversation history into isolated namespaces so department voices stay distinct instead of averaging into a committee tone, and reached 97% estimate-lookup coverage verified end to end against a real customer's quoted systems and prices. Designed and shipped in seven days with zero production impact.

Finance and campaign automation. Nightly invoice reconciliation against QuickBooks Online, deterministic invoice drafting from a versioned rules registry behind a fail-closed validation gate, and self-serve email campaigns from a brand-locked template behind a required human approval gate.

Reliability engineering. This is the part most automation work skips. After finding four scheduled jobs silently dead for up to 17 days, I rebuilt the job board around a single principle: health is proven by fresh output, never by an exit code. Watchdogs that date every artifact, completion markers written after work returns rather than at startup, explicit timeouts on every unattended network call, and staggered triggers to survive wake-from-sleep contention.

My own work, shipped and in use.

Members portal for a global-missions organization. Two audiences behind one login, with server-side geolocation and VPN or proxy flagging on every sign-in so leadership has an honest picture of the field while protecting workers in dangerous places. Never blocks, only flags and scores. Astro on Cloudflare Pages Functions with D1, KV, and R2, first-party auth with httpOnly session cookies, no third-party auth SaaS. Includes a member-ID module with an 8-character Crockford base32 serial plus a mod-37 check character, chosen because 37 is prime and coprime with base 32, so it detects every single-character substitution and every adjacent transposition. Both properties are asserted exhaustively in the test suite. tsc --noEmit clean, 53 acceptance tests passing.

Personal finance PWA on a serverless backend. A single-file progressive web app on Google Cloud Storage reading and writing a Google Sheet through a secret-protected Apps Script webhook. PIN unlock with AES-256-GCM key sealing in the Web Crypto API, WebAuthn PRF biometric unlock, on-device notification scheduling with quiet hours and rate caps, and a documented replication runbook for handing a fresh instance to another owner. Running cost approximately zero.

Meeting capture and transcription pipeline. An Android PWA records in-person meetings and syncs privately to a laptop watcher that transcribes with faster-whisper large-v3 on GPU and returns a structured note to the phone. Nothing leaves the owner's devices.

Shadow-and-learn knowledge extraction. A screen, OCR, and call-audio recorder that captures a dispatcher's shift, plus an analyzer that turns the session into an explicit scheduling playbook. Local-only by default, with an explicit and scrubbed cloud opt-in.

Remote build console. A Cloudflare Worker and D1 queue that dispatches headless AI build tasks to a laptop runner from a phone, with token-checked actions, per-project routing, local-only logs, and status notifications.

Scheduled task desk. A Cloudflare Worker that runs every weekday morning, reads a whole project workspace, finds every task that is drifting, and posts one message per person into Discord, each tagged in their own message with their own list, status icons, and a plain-English note on what the task needs from them. Roles are resolved by name at run time, so no numeric IDs are stored anywhere. Live since August 2026.

Languages Python, JavaScript, TypeScript, HTML/CSS

AI Anthropic Claude, OpenAI embeddings, retrieval-augmented generation on private corpora, per-namespace voice modeling, prompt engineering, human-in-the-loop design with escalation guardrails, prompt-injection review

Automation n8n, Cloudflare Workers, Google Apps Script, Windows Task Scheduler, webhook architectures, job health monitoring

Web Astro, Cloudflare Pages Functions, progressive web apps and service workers, Chrome extensions (MV3), embedded panels inside third-party tools

Data Pinecone vector search, MongoDB, Cloudflare D1, KV and R2, Google Sheets, SQLite, JSON pipelines

Applied cryptography AES-256-GCM, PBKDF2-HMAC-SHA256, WebAuthn PRF, first-party session auth with httpOnly cookies

Platforms Linux VPS operations, Google Cloud Storage, Cloudflare Pages, reverse proxy and TLS

Integrations QuickBooks Online, HubSpot, RingCentral, Mailchimp, GoHighLevel, FieldPulse, Salesmsg, Plutio, Discord, Google Distance Matrix

Operations EOS Integrator, Lean Six Sigma Green Belt, SOP architecture, offshore team leadership, M&A operational readiness and post-acquisition integration

I build things that report on themselves. Every system ships with its own observability: a version header, a UI footer, server-side telemetry. If a build needs someone to confirm it is working, it is not finished.

I treat a silent failure as worse than a loud one. A job that cannot start writes no log, and no log looks exactly like nothing to report.

I draw the boundary around the model before I write the prompt. A language model names things and words things. It does not touch a total, a tax decision, or an account, and it does not send anything to a customer without a human in the path.

• Teaching AI Fluency, Anthropic, September 2026
• AI Fluency for Builders, Anthropic, August 2026
• Lean Six Sigma Green Belt, Six Sigma Academy Amsterdam
• EOS Integrator
• Ateneo de Davao University, graduate

Based in Davao City, Philippines. Working with clients across US and AU hours.

Available for AI Director and Head of AI roles, and for fractional Chief AI Officer and CTO engagements.

Reach me at consultwithpatrick@gmail.com