Circuit boards, then code, then client conversations. Ten years on one thread. Scroll to follow it.
photo · me
photo · university days
2016
Circuits & curiosity
Electronic and Computer System Engineering at Silpakorn University. I learned to make things sense, count and decide, and wrote my thesis on letting genetic algorithms tune deep networks.
B.Eng., second-class honours
Fujikura scholarship, 2019
Thesis: Evolutionary Computation for Deep Learning
photo · the factory years
2020
First lines of code
System Analyst and Developer at Fujikura. Xamarin, .NET and Oracle on a factory floor: inventory scanners, device trackers, a smart-factory dashboard, and a visitor system built during COVID.
Fujikura Electronics (Thailand) · Jul 2020 – Dec 2022
Web, Android and UWP apps
Trainer for OutSystems and API integration
photo · Rising Star Award 2023
2023
Finding my platform
Senior OutSystems Developer at iPassion. Tech Lead on one project, Business Analyst on the next. I moved enterprise apps from Traditional to Reactive and learned Agile the practical way.
iPassion · Jan 2023 – Dec 2024
Tech Lead, BA and SA roles
Associate Tech Lead · Architecture Specialist
photo · training the team
2025
Leading the build
OutSystems Technical Lead and Solution Delivery. I ran delivery end to end, from the first requirement to go-live, and trained developers, interns and customers so new people could ship sooner.
iPassion · Jan 2025 – Mar 2026
Full SDLC lead, trainer and mentor
First presales: assessments, estimates, demos
photo · a client presentation
2026
Talking to people about problems
Technical Presale and Consultant. I stopped being just a developer and became the person who listens first. Twelve solutions designed this year, many proven with a PoC before the proposal was written.
iPassion · Jan 2026 – now
Proposals, estimates, architecture, PoCs
Computer vision, document AI, enterprise AI, PWA
The pinboard
Work I can talk about
Client names are left out. Each card opens the problem, the approach and what it took.
Solutioning · 2026
Presales: discovery, architecture, estimates, PoCs and proposals.
Agriculture · seed R&D
Field Inspector
Measure every seed with a phone camera, in the field instead of at a bench.
Problem
Seed length, width and grade were measured with calipers and typed into Excel. Measuring happened at a desk, took minutes per tray, and field photos, notes and lab sheets lived apart.
Approach
A phone-native inspector app: calibrated capture with LiDAR or an ArUco marker, an ROI overlay, on-device segmentation, per-seed metadata with GPS, an offline-first sync queue and CSV export.
Result
Model refinement raised precision from 0.67 to 0.83 on the evaluation set.
Count sacks moving between people, containers and conveyors, live.
Problem
Transfers were counted from recorded videos in a phone app. Every step needed a person, and results arrived after the fact.
Approach
Edge detectors at container corners and conveyors count sacks crossing a line, a central server aggregates many devices, and a dashboard shows counts in real time. Built for day, night and dust.
Result
The lightweight model reached mAP50 0.90 at 3.1 ms per image on the edge device.
What it took
YOLO11 on Raspberry Pi 5 + Hailo-8LONNX / HEFNode-REDPostgreSQLWebSocketPortainer
From Bill of Lading and invoice to an approved, QR-verifiable certificate.
Problem
Every fumigation certificate was retyped from two source documents. One typo meant a rejection at customs, and forged certificates carried the company's name.
Approach
One extraction core reads the B/L and invoice, enforces matching keys (container, invoice and PO numbers), drafts the certificate for approval and stamps a QR code anyone can verify. Manual and automated modes share the same core.
Result
The PoC was working before the proposal was written.
Reservations, lots, delivery and after-care in one Thai-language PWA.
Problem
Bookings lived in notebooks and spreadsheets, ID details were copied by hand, lot assignment happened verbally, and sales were re-keyed into the ERP.
Approach
A mobile-first PWA in four selectable phases: sales and capacity control, loyalty, production and QC, and after-sale care. Business rules are enforced as hard gates in the API, with ERP batch sync.
The counter could not see real stock, every sale was re-keyed into the ERP, mall reports were typed by hand, and herbal extracts were deducted manually.
Approach
An iPad storefront and cashier PWA that syncs to the existing ERP: ratio-based extract deduction, FIFO lots, automatic mall GP split, split payment, real-time multi-branch stock and permission-gated discounts.
Result
One core, configured per branch, for a growing chain.
What it took
PWA on iPad and desktopREST servicesERP integrationPromptPay / 2C2P
Read an OutSystems ODC app's logic out automatically before rewriting it.
Problem
ODC has no export of application logic. Without one, rewriting a system means guessing what it does, and nobody can estimate a guess.
Approach
A PoC that decodes each app's flows, links and data model, lists every gap with its location, and turns the revamp into a counted translation job. Migration plan: .NET with React/Next.js on Azure, piloted on two stable apps.
Result
3,218 of 3,219 flow nodes recovered; 99.1% ready to code without opening ODC; about 3 seconds per app.
Claims and advances in one place, with receipts read by AI.
Problem
The TOR asked for one hub for expense claims and cash advances, AI OCR with at least 90% accuracy, and integration with SAP.
Approach
Journey and integration maps, a design system and clickable prototype, and an OCR PoC that tested three engines against the TOR's AI clauses and priced each option.
Result
A clear vendor comparison against every AI clause in the TOR, with the cost of each option side by side.
What it took
Azure Document IntelligenceABBYY VantageClearReadSAP S/4HANA integrationResponsive web
Versions, approvals and audit trail instead of shared folders and email.
Problem
Documents sat in shared drives, approvals ran through email, nobody could tell which revision was current, and there was no audit trail.
Approach
Two options side by side: a custom offline-capable PWA with a multi-stage workflow engine, role-based access, wiki, issue tracker and share links, or an Alfresco ECM platform.
Result
A working demo for each option, so the client could choose by seeing.
Four social projects across schools nationwide kept scores, forms and photos in Excel, PDF, LINE and paper. One sponsor report took months to assemble.
Approach
One platform with four roles: teachers record and see only their school, the foundation approves and reports, IT manages access, and executives read KPI dashboards and an impact map. Priced module by module.
Result
Delivery plans and priced options the client could compare line by line.
What it took
Web platformRole-based accessKPI dashboardsImpact map
Transporters collect parts from many suppliers on one route, but the distance being paid and the run actually driven were never in the same system.
Approach
Approved routes become the master for payment, GPS and odometer readings confirm each run, both billing models are handled (distance per route and a province price matrix), and approved amounts flow to the ERP as purchase requests with a full audit trail.
Result
Every approval, route and payment in one auditable flow.
What it took
Web platformGPS & mileage verificationERP integrationApproval workflow
One AI portal for a whole organisation, under its own domain.
Problem
The TOR asked for multi-LLM access, custom agents, RBAC with SSO and very high token volumes. Commercial platform licences alone would have exceeded the budget.
Approach
Compared five architectures, then proposed a self-hosted open-source stack: a LibreChat front end, Dify as the engine, LiteLLM as the model gateway and Langfuse for observability.
Result
Full TOR coverage with zero software licence cost.
Enterprise AI on the customer's own servers, with the budget under control.
Problem
Teams buy AI tools separately, costs climb without anyone noticing, company data leaves the building, and the organisation depends on one vendor.
Approach
One portal installed on the customer's infrastructure: a single AI budget allocated to departments and people with automatic cut-off, free choice of models (OpenAI, Gemini, Claude, Grok), a Thai-language knowledge base with RAG, and privacy by design.
Result
Budget, access, data and long-term cost controlled in one place.
A timesheet PDF becomes a checked, ready-to-send invoice.
Problem
Each monthly bill meant reading timesheets by hand, matching working days and leave, and typing the invoice. Slow, and easy to get wrong.
Approach
Six steps: receive the approved timesheet, extract it with document AI, check days against the calendar, approve through LINE as a human gate, draft the invoice in the accounting system, and draft the client email. Every step is logged.
Result
A billing cycle with one human decision instead of a manual chain.
What it took
n8nABBYY VantageClearRead AILINE approvalAccounting system APIAudit log
How to prompt, DESIGN.md, Google Stitch and Claude Design, with live demos.
Approach
Prompting principles, why a DESIGN.md keeps AI output consistent, and live walkthroughs of building UI with Google Stitch and Claude Design, from prompt to export.
B.Eng. Electronic and Computer System EngineeringSilpakorn University, Sanam Chandra Palace · 2016 – 2020Second-class honours · Thesis: Applications of Evolutionary Computation for Deep Learning
Mathematics and Science programmePhothawatthanasenee School · 2013 – 2015Academic excellence
Now
Still building
If you have a problem that needs a line drawn through it, from the first conversation to a working system, write to me.
