kuanyu.dev — forward-deployed AI engineer
Hi, I'm Kuan. I turn messy work into systems that run.
A forward-deployed AI engineer who designs and architects workflows. I sit with a real, messy process, work out the version that actually happens, and build the first one that ships — keeping a human hand wherever it matters.
Here's what I've built, sorted by the problem it solves. ↓
The work
Regulated tax work is buried in PDFs nobody wants to read.
Corporate tax in Singapore is buried in messy PDFs and rules no one wants to read.
Codex/GPT-native, local-first public demo/runtime port of the AITax Singapore corporate-tax computation pipeline.
Experts rebuild the same deliverable from scratch, every single time.
A Big-4-trained consultant rebuilds every incentive proposal from scratch.
Same-day delivery for a consumer-dental company whose dev is outsourced: a Custom GPT running the client PM's requirements skill with a GET-only Bitbucket Action (scoped read-only token), so any employee can paste a vague stakeholder request and get an evidence-backed classification — 'already supported, here is the file' vs 'new feature, here is the PRD' — before a ticket reaches the metered dev vendor.
Find, enrich, reach, follow up — the outbound grind run by systems, not people.
High-value healthcare/genomics engagement: a Patient Clinic Finder to coordinate IVF clinics with overseas genomics labs.
Property agents sit on hundreds of dead WhatsApp leads they never recontact.
Paid sales-outreach automation package.
Writing a genuinely personalized B2B email per company doesn't scale by hand.
On a live sales call there's no time to think of the right next line.
Reusable scraper framework: per-site adapters with crawling, rate-limiting, contact extraction and JSONL export.
Autonomous 24/7 sales orchestration running a tmux-based three-tier AI agent hierarchy from discovery to close.
Cold-outreach pipelines fail mid-run — and you can't tell which step broke or restart it cleanly.
LinkedIn crawler targeting recently-laid-off Singapore professionals for personalized DM outreach.
Communities coordinate everything by hand, in group chats nobody can search.
Staying genuinely in touch stops scaling past a few dozen people.
Great coaching lives in one expert's head — one athlete at a time.
AI running coach turning Jack Daniels' VDOT plans into per-run coaching delivered to a Garmin watch.
Client workflow engagement nested inside AI Workflow School (workflowlab-sg customers).
The feed is mostly noise — the few posts worth reading get buried.
Your own health data is scattered across clinics, wearables, and PDFs that never talk to each other.
Your medical history is scattered across clinics, labs, and apps that never talk to each other
A diabetic family member's signals were spread across a glucose monitor, a wearable, and lab PDFs that nobody cross-referenced
now · where I'm pulling next
The threads I'm actively working into something bigger
about · the short story
I design and architect workflows — and the technology around them
I've spent my time operating and building inside real workflows — doing the actual work, not just specifying it from the outside. That hands-on experience is where the real understanding comes from: you only see where a process breaks when you've lived it.
Then I take a step back and look at the whole thing. From first principles, I ask whether the work could be structured differently — and genuinely better — instead of just automating the mess exactly as it is.
So the job is the same under every title: decompose a messy process into a system — state made explicit, tools that do the steps, a person checking what matters, and logs so it keeps improving. Architect first, automate second.
say hi
Got a messy workflow worth fixing?
I'm open to forward-deployed roles and scoped client builds — especially in GTM, regulated work, and health. If something here looks close to your problem, send it over.