Automation · RPA · Agentic AI

Automation flows, end to end.

I turn manual, email-driven operations into reliable automated pipelines — and design multi-layer AI agents that perceive, decide and act on their own. This is a working library of flows I've architected: the problem, the orchestration, and what it changed.

Power Automate n8n Claude API Azure Document Intelligence Playwright Office Scripts Python

A note on confidentiality. Client names, source code (Office Scripts), field-level mappings and internal paths are deliberately withheld. What you see here is architecture and sanitized, redacted samples — enough to judge the engineering, nothing that exposes an employer or customer.

Featured build

01 / flagship
Independent side project

Multi-Layer AI Agentic Platform

A self-directed agent that perceives events, reasons, and acts across the web with no human in the loop — and recovers from its own failures. Built to run 24/7 as a service.

The idea

Most "automation" is a fixed script. This is an agent: it observes a live environment, asks an LLM what to do next, carries the action out in a real browser, checks its own work, and retries when something breaks — looping until the goal is met.

Why it's hard

The web changes under you. The win is the orchestration + guardrail layer (n8n) that keeps a long-running agent stable: idempotent steps, retries, and a feedback path so a failed action re-enters reasoning instead of crashing the run.

What it shows

End-to-end agentic design — API-driven reasoning, workflow orchestration, browser RPA, monitoring and alerting — the exact ingredients of a production internal automation.

Stack
n8nClaude APIPlaywrightPythonTelegram Bot APIWindows service
Download
Architecture sheet01 · PDF · 1 page

Automation flows

02 / production patterns
Document AI · full loop

AI Document-Extraction Pipeline

Customer emails a PDF; a standardized template comes out the other end — no human re-keying.

The whole loop runs in Power Automate. A per-line arithmetic checksum catches OCR mis-reads before they hit the template — the detail that makes it trustworthy. A revision branch diffs a re-sent document against the last version.

Power AutomateAzure Doc IntelligenceOffice ScriptsSharePointTeams
Architecture sheet02 · PDF Redacted flow skeletonJSON · structure only
Data pipeline · event-driven

Event-Driven KPI Pipeline

Turns scattered milestone emails for a key account into a live, time-stamped operations dashboard.

Three parallel event flows (issuance, revision request, revision response) feed one structured table. Event-driven, so metrics update the moment an email lands — no batch jobs, no polling.

Power AutomateExcel OnlineSharePointKPI reporting
Architecture sheet03 · PDF
LLM integration · internal assistant

Claude API Email Triage

An always-on assistant that reads inbound mail and posts a Thai-language digest to the team's chat.

Replaced a discontinued paid AI service at zero added licence cost by calling the Claude API directly from the flow — a practical internal AI tool the operations team actually uses daily.

Power AutomateClaude APIMicrosoft Teams
How these get built

One method, many customers

The extraction pipeline isn't rebuilt per client — each customer is a profile on a shared method.

Adding a customer means writing a small profile, not a new app. That's what keeps a 5-person team able to serve many stakeholders — scalable by design, documented with SOPs so others can run it.

Reusable skillsSOPsPer-customer profiles

What I bring

03 / capability
01

Workflow orchestration

Power Automate & n8n — long-running, idempotent, self-healing flows.

02

RPA

Playwright browser automation where no API exists.

03

LLM integration

Claude / GPT APIs wired straight into operational flows.

04

Document AI

Azure Document Intelligence with arithmetic validation.

05

Reliability

Monitoring, retries, troubleshooting, SOPs for handover.