I'm Tienwito Karnthanat — I build automation flows end to end: RPA and document AI to remove manual re-keying, LLMs wired straight into the loop, and multi-layer AI agents that decide and act on their own.
Most automation fails because nobody measured the problem. I start with time-and-motion analysis — breaking a workflow into tasks, attaching real handling times and volumes, and converting it into FTE workload.
I work inside large, compliance-heavy organisations where data is sensitive and change is slow. My job is to make automation defensible to management and safe for operations.
I specialise in document-heavy workflows — invoices, customs declarations, certificates — where AI extraction meets real compliance stakes.
Manual re-keying from scattered PDF layouts into customs declarations. High volume, high error cost, zero scalability.
Azure Document Intelligence pipeline parsing data directly into Excel templates via Office Scripts. Separate logic and extraction layers.
Repeatable Python pipeline merging thousands of OCR'd rows, with a human-in-the-loop QC checklist. 100% accuracy required for customs compliance.
A billing flow for a global e-commerce account spanned three disconnected systems. Solved the cross-origin barrier with a lightweight proxy so data moves automatically.
Real-time and backfill flows parsing alert emails into SharePoint Excel. Replaced a discontinued AI feature with a direct Claude API call, packaged as a reusable team skill.
A layered AI-agent architecture for a solo online operation: browser automation for repetitive web steps, an LLM layer to classify freeform replies, success/exception branching with retries, and remote control via a chat bot. Independent R&D in agentic workflow design.
Want the architecture behind these? The full flow library has diagrams and downloadable, redacted samples.
Explore the automation flow library arrow_forwardBetween projects I build and ship small tools — it's how I stay sharp and test ideas before they reach production work.
Packaging components so teams can apply LLM logic without rebuilding the architecture every time.
Scripting the custom connections that off-the-shelf tools can't handle, from data cleanup to API orchestration.
Replacing discontinued no-code AI features with direct API calls (Claude/GPT) in live business flows.
Designing multi-layer AI agent architectures that decompose complex operations into manageable stages.