Independent contributor — AI & Automation

I find the gap, build the system, ship it before anyone asks.

I build AI and automation systems embedded inside operations — customer support, hardware quality, workflow routing, and the bridges between enterprise platforms that weren't designed to talk to each other. Production systems, not proofs of concept. Every project below shipped, measured outcomes, and is still running.

6+production systems
$630K+annual value created
95%+CSAT
10+ yrsoperations
See selected work ↓ Get in touch
Selected Work

Six systems. One operating pattern.

Find the broken process. Build the fix. Prove it in production. Document the wins. Move on. Every project below shipped without formal headcount or budget approval and earned its place by demonstrating measurable value before being asked.

CARA

Autonomous customer support AI
In production
Problem

Support team capacity was the gate on response speed and quality. We wanted more control over the customer-facing voice end-to-end — deeper KB integration, multi-step reasoning, quality scoring per response — rather than evolving at a vendor's roadmap pace.

Approach

Built autonomous support AI from scratch, integrated with the existing ticketing platform. Domain-specific knowledge layer, multi-step reasoning, per-response quality scoring. Shipped without formal headcount approval.

Outcome

8–12 FTE equivalent workload at 95%+ CSAT. ~$500K/year in capacity creation. Successor (CARA 3.0) brings AI infrastructure in-house at ~$4K/year in direct API spend.

Claude API Zendesk API Node.js Custom KB

QC Dashboard

Hardware quality intelligence
Live
Problem

~35K RMA records sitting in a Bitable with no real-time visibility for hardware engineering, product, or QC. Quality spikes were caught weeks after the fact in monthly review meetings.

Approach

Built dashboard with incremental Bitable sync (135× faster than full pulls), multi-device trend overlays, severity-scored spike alerts. Failure-rate math wired and ready for activation-count integration.

Outcome

Quality data democratized across engineering, product, QC. Per-device drill-downs let engineers find SKU-specific patterns monthly reports miss. Spike alerts catch issues weeks earlier.

Express SQLite Bitable API Chart.js
Interactive demo — sanitized data, real UI →

Lark Mail Router

Inbound triage at scale
Live
Problem

Single inbound mailbox served 22 distinct audiences (regional reps, partnerships, customer support, B2B distributors). Broken IMAP pipeline. Multilingual. Mis-routes daily.

Approach

Built classifier on Claude Haiku. Multi-rule routing table with category overrides (customer-support topics always route to global support regardless of language). Dashboard captures human corrections as few-shot training examples.

Outcome

Inbox triaged automatically with self-improving feedback loop. Support-topic policy enforced across 22 categories and 8+ languages.

Claude Haiku Feishu Mail API Node.js PM2

RELAY

Custom support platform
Built · on hold
Problem

Zendesk got the support operation to a working baseline. As we matured, we wanted a platform that fit our workflows exactly — custom routing, deeper CARA integration, the ability to ship UI changes on our own schedule. Owning the customer experience meant building our own.

Approach

Built complete custom platform from scratch — 48 routes, 19 background workers, 25+ services, 55+ frontend components, 61 DB migrations. Three-panel CARA simulator, Zendesk article import, full ticket lifecycle.

Outcome

Working production-ready platform. Held pending an organizational chat-strategy decision — knowing when to wait is part of shipping. Activatable the moment alignment lands.

Node.js / Express PostgreSQL React Tailwind

SENTINEL

Brand protection monitoring
Live
Problem

Fraudulent lookalike sites impersonating the brand with similar domains, stealing customers and trust. No detection in place — discovered only when customers complained.

Approach

Python scanning modules — DNS twist analysis, Certificate Transparency log scanning, search-engine crawling. Next.js dashboard. Documented takedown workflow.

Outcome

Lookalike domains caught in CT logs within hours of registration. Reactive customer-complaint problem converted into proactive defense.

Python Next.js CT logs DNS twist

AI Playbook & COROSclaw

Adoption & enablement program
Rolled out
Problem

AI tools only work if the team alongside them does. Without a documented adoption framework and hands-on training, even strong systems gather dust.

Approach

Co-authored enterprise-wide AI Playbook with executive leadership. Deployed an internal Claude interface (built on the open-source OpenClaw foundation). Ran intro training for ~20 staff with full enablement package: deck, agenda, quick reference card, power-user guide.

Outcome

Adoption framework live across the company. Level 2 power-user training queued. Pattern replicates: build the system, train the humans, document the wins, measure.

Claude OpenClaw Gamma Feishu Docs

Try the interactive QC Dashboard demo linked above — same UI, sanitized data. For the others, ask and I’ll walk you through them in detail.

Get in touch →
Background

Operator first. AI builder second.

Ten years at Apple, two at COROS, one founded company on the side. The pattern across all of it is the same: see the gap, build the thing, prove it works, move on.

I started at Apple support in 2014, advanced through Special Projects, and ultimately served as Team Lead Admin on the AMR Commitments Team. That decade shaped how I think about customer experience: not as a department, but as the discipline you build everything else around.

At COROS I joined as Head of Customer Support in October 2023, where I built and led a 32-person global team across the US, EU/UK, China, and the Philippines. By mid-2025 the work had outgrown people management — the highest-leverage thing I could do was build AI and automation systems that scaled the team's capacity rather than its headcount.

So I designed my way out of that role. I helped hire and train a successor for the people-management side, and pivoted into a full-time AI & Automation position. I'm an individual contributor now, which is exactly the shape of work that lets me ship the most.

The AI work at COROS wasn't assigned. I saw the gap, taught myself what I needed, and built working systems that proved their value before they were ever official.

CARA — the production AI handling 8–12 FTE-equivalent of customer support at 95%+ CSAT — wasn't on anyone's roadmap. Neither was RELAY, the custom support platform built to give us full ownership of the support stack. Neither was SENTINEL, the brand-protection monitor that watches for fraudulent COROS lookalikes. Every one of them got built and proven before it had a budget line.

But the systems are half the story. I co-authored an enterprise-wide AI Playbook with executive leadership — a working adoption framework, not a PDF that collects dust. I run ongoing enablement for the support and product orgs because the systems are only as good as the team operating alongside them.

That's the pattern: find the real problem, build the thing that solves it, enable the people around it, measure everything, move on to the next one.

Build log

What I've been building this year.

Reverse-chronological log of work that landed in production or made meaningful progress. Filtered to the last 12 months. Click a status to filter.

May 12, 2026 Live
Lark Mail Router — sales@ auto-classification
Replaced a broken n8n IMAP pipeline with a purpose-built classifier. Polls the COROS sales mailbox via Feishu UAT, classifies each thread with Claude Haiku into one of 22 routes (regional reps, partnerships, API, customer support, auto-replies, spam, etc.), and forwards. Multi-user dashboard with feedback corrections that train the classifier as few-shot examples. Critical routing rule: customer-support topics override language and always go to global support.
See case study ↓
May 12, 2026 Live
QC Dashboard — hardware quality intelligence
Production dashboard for hardware engineering and QC. Live feed from a 35K-record Feishu Bitable, incremental sync (135× faster than full fetch via search-since-cursor), multi-device trend overlay, spike alerts with severity scoring, real-time per-device defect breakdowns. Failure-rate math is wired and waiting on monthly activation counts.
qc.infinitusdigitalllc.com · internal access
May 5, 2026 Shipped
COROSclaw — internal Claude UI rollout
Led the 60-minute intro training for ~20 COROS employees on COROSclaw, our internal Claude-powered interface. Built the full enablement package: Gamma deck, agenda, Quick Reference Card, Power User Guide — all delivered as both Word docs and native Feishu documents. Level 2 Power User Guide session is queued.
Apr 2026 Live
Feishu / Lark MCP infrastructure
Dual-server MCP architecture (tenant token bot + user-context wrapper) connecting Claude directly to internal COROS systems. Enables natural-language operations on Bitable, calendar, tasks, messaging, and docs from any Claude surface. Quietly powerful — turns natural-language requests into structured API calls without leaving the chat.
Apr 2026 Live
SENTINEL — brand protection monitor
Watches for fraudulent COROS lookalike sites using DNS twist analysis + Certificate Transparency log scanning + DuckDuckGo crawler. Python scanning modules + Next.js dashboard, PM2-managed on the production droplet. Takedown workflow documented; Feishu announcement integration on the roadmap.
Brand protection · internal tool
Mar 2026 Live
VOICE — customer feedback intelligence
Fider-based feedback platform with NLP analysis layered on top. Pulls 12K+ support tickets, categorizes by theme, surfaces emerging trends with sentiment scoring. Replaces what used to be an 8-hour weekly manual analysis — now updates in real time.
Customer feedback intelligence · internal tool
Mar 2026 Live
SCOUT — competitive monitoring
Continuous monitoring of Garmin, Suunto, Polar, and Wahoo product pages and content for launches, pricing changes, and positioning shifts. Quiet daily output that informs sales, marketing, and product teams without anyone having to remember to look.
Feb 2026 Built · on hold
RELAY — custom support platform
Custom support platform built from scratch — Node/Express, 48 routes, 19 background workers, 25+ services, 55+ frontend components, 61 DB migrations. Three-panel CARA simulator interface, Zendesk article import, full ticket lifecycle. Built, deployed, working. On hold pending an organizational decision on the chat strategy — being patient is the right move here.
2025-Q4 → ongoing In development
CARA 3.0 — next-gen agentic architecture
Successor to CARA with multi-step reasoning, real-time quality scoring, and per-channel personality consistency. Built on Anthropic's API directly — projected operating cost ~$2,400–4,800/year for equivalent volume, bringing AI infrastructure fully in-house. Active development, not yet launched.
2025 In production
CARA — autonomous customer support AI
The original. Handles the equivalent workload of 8–12 full-time employees at 95%+ CSAT, 24/7. Built without formal headcount approval — proved its value in production before becoming the anchor of the AI program. ~$500K/year in capacity creation.
How I work

Four convictions behind every system.

These aren't slogans. They're the operating rules that show up in the code, the architecture decisions, and the conversations with stakeholders.

01
Prove it before you pitch it
Working systems beat decks. CARA, RELAY, SENTINEL, the QC Dashboard — all of them ran in production before they had budget lines. The fastest way to get something funded is to make it irreplaceable while no one was paying attention.
02
AI amplifies people, it doesn't replace them
The 8–12 FTE capacity that CARA handles isn't headcount that got eliminated — it's capacity that lets the team focus on the harder, more interesting problems. The AI Playbook is the other half of that work: making sure the team can actually operate alongside the systems.
03
Off-the-shelf has a ceiling
Vendor platforms get teams to a working baseline quickly, and that has real value. But there comes a point where you want more control over the customer experience — deeper integration, custom workflows, the ability to move strategy and execution in sync. That's when you build. RELAY and CARA 3.0 are both expressions of that moment.
04
Transformation fails without enablement
Every system here ships with documentation, training, and ongoing support for the humans next to it. The hardest part of AI transformation isn't the AI — it's changing how people work alongside it. Both halves get my time, intentionally.
Get in touch

Open to the right conversation.

If you're looking for someone who finds the gap, builds the system, and ships it before anyone has to ask — let's talk. I'm interested in IC and senior IC roles that combine agentic AI work with real operational stakes. Also available for select consulting engagements through Infinitus Digital.

Click to contact →
Location Washington State · Remote
Availability Open to conversations