Application Development Engineer at Delta Electronics. Go, Python, log platforms. Open source contributor to mcp-go.

github

Ekdanai Kummee • Peerlist

Containerized e-commerce platform with real-time stock and order handling, automated delivery services, and an admin back office.<ul><li>Go backend with Redis for carts and sessions, and Redis distributed locks so two buyers can't take the last item at once</li><li>Next.js frontend with Discord OAuth2 sign-in</li><li>MariaDB with transaction isolation tuned to keep checkout consistent</li><li>Docker Compose deployment, with Kubernetes Horizontal Pod Autoscaler for scaling</li><li>Cloudflare WAF, API rate limiting, activity logs, and HMAC-signed payment validation</li></ul>

Raven Webmarket E-Commerce

RoomedIn is a web application built to solve the communication gap between housekeeping, reception staff, and owners at small hotels and guesthouses, who typically rely on walkie-talkies or spreadsheets to track room status.<ul><li>Real-time room status board, every change reflects instantly across all devices</li><li>Role-based access (owner, admin, reception, housekeeping)</li><li>Guests order food/services and chat with staff by scanning a QR code, no app install or login required</li><li>SOS button that instantly alerts all staff in an emergency</li><li>Finance dashboard summarizing income, expenses, and staff activity</li><li>Multi-hotel support on a single account, switch between properties from one place</li><li>Platform admin system for managing account access, banning/deleting hotels, sending notices to owners, plus a support ticket and live chat system</li><li>Bilingual (Thai/English), with a 30-day free trial before a monthly subscription</li></ul>

RoomedIn: Real-Time Room Status

Built the log analysis platform for the factory's MES. It reads the logs of the API gateway, routing, SMT machine and IoT aggregator services into OpenSearch, finds the issues, explains them with an AI assistant, and alerts Teams while production is still affected.<ul><li>Go ingest path with only the standard library: the collector downloads just the new bytes of each log with HTTP Range requests, auto-scaled workers run the parsers, and an importer writes daily indices</li><li>Logs and alerts run one to two minutes behind by reading the live file, and each record id is a hash of its line, so re-imports never duplicate</li><li>Parsers split gateway errors into exception, root cause, class and method, show SQL to IT only, and flag out-of-limit measurements and template values sent instead of real machine data</li><li>19 alert rules in three levels, judged on the logs' own clock, so only incidents still going in the newest minute are sent</li><li>Seven Teams destinations as Adaptive Cards or channel email through the company mail server, through an outbox with retries</li><li>Issues page and AI answers whose links open exactly the records they talk about</li><li>External API with read, tune and admin keys for other AI tools, with every change logged and undoable</li><li>GitLab CI builds only changed images and deploys the API blue/green behind nginx. About 300 Python tests, Go unit tests and 140 Vitest tests run in CI</li></ul>

AI OpenSearch: Factory Log Analytics

Solo-built, open-source identity provider written in Go, from an empty repo to a tagged v1.0.0 release. Ships as a single static binary (no cgo) that embeds a Next.js admin console via go:embed, so deployment on both Linux and Windows is just "copy one file," whether as a native service or a container.<ul><li>RBAC: users through groups (with hierarchy) through roles through permissions, cached and invalidated on write</li><li>Real OIDC provider: authorization_code with mandatory PKCE, refresh_token, client_credentials for service-to-service auth, JWKS, discovery. Works with GitLab, Harbor, Rancher, Grafana, Jenkins, Vault, NiFi, MinIO, and anything else that speaks OpenID Connect</li><li>WebAuthn/FIDO2 and TOTP for MFA</li><li>Pluggable database layer (Postgres, MySQL, MSSQL, SQLite) through dialect-aware SQL, no ORM</li><li>Real clustering: DB-backed leader election for background jobs, Redis pub/sub fan-out for realtime WebSocket updates across instances</li><li>Security-first onboarding: every account gets one server-configured default password and is forced to change it on first login. No endpoint anywhere accepts a caller-chosen password for someone else, and none ever returns one</li><li>Permission-aware admin UI: the frontend hides whatever a signed-in user isn't allowed to do instead of rendering it and letting the API reject it after the fact</li></ul>

IdpForge: Self-Hosted SSO in Go

Built the CI/CD platform for the factory's 24/7 MES portal and was the first on the Thai team to put this architecture together. Other application teams and the server team are now adopting it, with me as the point of contact.<ul><li>GitLab CI/CD pipelines for the MES portal frontend (React/Vite) and API (ASP.NET Core)</li><li>Images pushed to Harbor and deployed to a Rancher-managed Kubernetes cluster inside an air-gapped network</li><li>Horizontal Pod Autoscaler for traffic spikes at shift changes</li><li>Vulnerability scanning as a gate before production deploys</li><li>Rolling updates with liveness and readiness probes, so deploys cause no downtime</li><li>Resource limits and health checks across every deployment</li><li>Presented the architecture to the whole IT department</li></ul>

MES CI/CD Platform on Kubernetes

Designed the data flows that pull production data out of the MES and deliver it to downstream systems, running on the NiFi cluster and central Kafka platform I set up.<ul><li>NiFi reads from the Oracle MES databases over CDC/JDBC and produces raw events into Kafka</li><li>Spark Structured Streaming consumes the raw topics and writes cleaned, deduplicated gold topics for downstream apps</li><li>Airflow workflows schedule the Spark jobs every 15 to 30 minutes, with retries and run history</li></ul>

MES Data Pipeline (NiFi, Kafka, Spark)

Founded and ran a multiplayer roleplay server platform from scratch, scaled to 2,000+ concurrent players over several years. This is where most of my backend, database, and systems experience comes from.<ul><li>Designed every in-game interface and web dashboard in Figma before writing code</li><li>Built lightweight in-game UI in Svelte and TypeScript</li><li>Wrote the server systems in Lua and tuned them for heavy load</li><li>Designed the MariaDB schema and fixed slow queries that caused lag spikes</li><li>Audited and rewrote legacy code to close client-side exploits</li></ul>

FiveM Multiplayer Platform

Contributions to mcp-go, a Go SDK for the Model Context Protocol with 9,000+ stars.<ul><li>Fixed the streamable HTTP client losing the real HTTP status when a gateway, CDN or web framework answers with a JSON error body (merged)</li><li>Made a flaky SSE transport test deterministic</li><li>Helped triage issues already fixed on main</li></ul>

Open Source: mcp-go (Go SDK for MCP)

Full-stack production yield reporting and drill-down tool built solo for the QA team at Delta Electronics Thailand. Aggregates station-level pass/fail data across factories and models, with drill-down to individual repair records and Excel export. Fixed an N+1 query, parallelized cross-schema data fetching, added row virtualization for 10,000+ row tables, and traced a session-hydration race condition that had been silently breaking deep links app-wide. Deployed in Thailand with rollout planned for Taiwan.

Yield Tracker Report

Self-hosted bridge that connects any Kafka topic to any HTTP app with no Kafka client code, built solo from an empty repo. One Go binary, one YAML file, and a React console.<ul><li>At-least-once delivery with in-order commits, retries with backoff, per-key ordering, dead-letter and reject topics, and a circuit breaker</li><li>Flows as a DAG: conditions, data checks, fan-out to several apps, webhooks, and topics, with a separate line for every answer, reject, and failure</li><li>Ready-made steps for OpenSearch, Elasticsearch, NiFi, another Kafka cluster, Slack, Discord, and Teams</li><li>Cluster mode coordinated through Kafka itself: leader election, label-based placement, failover in about a second</li><li>Built-in MCP server and assistant that diagnose pipelines and draft config changes applied only after a confirmed preview</li><li>Live React Flow canvas where each dot is ten real messages</li><li>gosec, Semgrep, govulncheck, OSV-Scanner, Gitleaks, Trivy, and CodeQL on every push</li></ul>

ARK: Kafka-to-HTTP Bridge

Desktop app that turns YouTube Music into a live Discord Rich Presence status, plus a spinning-vinyl streaming overlay for OBS/Streamlabs. Built solo in Electron from an empty repo through 8 tagged releases in one sitting, with a full CI/CD pipeline: GitHub Actions builds and auto-publishes a Windows installer on every version tag, and the app checks and applies its own updates in the background via electron-updater.<ul><li>Reverse-engineered YouTube Music's undocumented player DOM to read now-playing state, and worked around its Trusted Types CSP (which blocks raw innerHTML) by registering a custom Trusted Types policy for the injected UI.</li><li>Diagnosed a Discord RPC bug where an unregistered Rich Presence image asset made Discord silently drop the whole activity payload, with no error over IPC. Found it by testing fields one at a time against a live Discord client instead of trusting the "success" response.</li><li>Worked around a Windows toolchain limitation in electron-builder (its icon-signing step needs Developer Mode) by writing a custom packaging step with the standalone rcedit tool, verifying the fix directly against the built executable's version resource.</li></ul>

Vinylcord

Web tool that traces production data errors and finds incorrect data configurations on the lines, a check that used to be done by hand.<ul><li>Pulls production data from Oracle and compares the MES schemas automatically with CTE-based queries, flagging each case as MATCH or MISMATCH</li><li>Cross-references live factory data from MS SQL Server and PostgreSQL</li><li>ASP.NET Core REST API that keeps the reporting logic apart from the UI</li><li>React and TypeScript UI that shows production line errors in a form the floor team can act on</li><li>Tuned SQL queries and stored procedures to cut report time and fix data latency</li></ul>

CTP Report System

Open-source Go library for type-safe HTTP APIs on net/http. One handler struct gives request binding, validation, an OpenAPI 3.1 document, and generated clients, all derived once at registration so the request path only runs precomputed plans.<ul><li>Binds path, query, header, cookie, form and body fields from struct tags</li><li>Generates a TypeScript client where each endpoint's declared errors form a typed union, plus Zod schemas with the same validation rules</li><li>Built-in docs page, a dev loop that rebuilds and regenerates on save, and a mock server</li><li>CI check that fails when the generated client is out of date</li></ul>

tack: Type-Safe HTTP APIs for Go

A code snippet manager built as two front-ends, a VS Code extension and an Electron desktop app, that read and write the same local data.json and stay in sync in real time via file watching (chokidar), with no server, no account, and no cloud involved. Built as an npm workspaces monorepo split into three packages: core (storage, CRUD, and file-watch sync), a VS Code extension, and an Electron app. The desktop app has a glass and blur visual style across three themes, generated from design tokens in the oklch color space, plus a VS Code Explorer style category tree with drag and drop reorder, a from scratch syntax highlighter, and full EN and TH support. The VS Code extension is the lighter surface, a sidebar tree grouped by language with click to copy, insert, or open. Distributed via electron-builder with GitHub Actions building and publishing installers on every tag push, plus in-app auto-update. Built July to August 2026, MIT-licensed, shipped at v0.6.0.

Snippet Manager

Built AI Center, the web portal that brings together every AI project the team has built in the company, plus AI Deltron and AI Monitor behind it.<ul><li>AI Deltron, the assistant that answers from company knowledge: retrieval over documents synced from SharePoint and Confluence and past ServiceDesk Plus (SDP) resolutions, embedded with Qwen models and reranked, with vision-language models for scanned documents and speech-to-text for audio</li><li>Knowledge built in stages: reviewed documents, data tables the model can call as tools, a document graph of attachments and versions, a change log, and evaluation runs scored per chat mode</li><li>AI Monitor, an admin console for every AI project: usage, chat logs, knowledge, prompts, evaluation, AI gateway use, services, and the audit log</li><li>Split from one FastAPI app into Python roles (api, worker, media, chat) and Go services (edge router, container scaling, an AI gateway proxy that shares the per-minute quota and queues instead of failing on 429, analytics), one undoable step at a time</li><li>Postgres with pgvector, MinIO, NATS JetStream, and a React 19 frontend that works from phone to desktop</li><li>Conversations logged to build fine-tuning datasets</li></ul>

AI Center, AI Deltron and AI Monitor

Set up and run a dedicated NiFi cluster as a central platform that IT and other teams build their own flows on, separate from the team's shared production clusters.<ul><li>High-availability NiFi cluster on RHEL, coordinated by a 3-node ZooKeeper ensemble for cluster state and failover</li><li>SSO login and access policies per team and per canvas area</li><li>Used as middleware for logging, alerting, data migration, and routing data across systems, including scripts for MinIO jobs</li><li>Integration and performance testing before handing it to other teams</li><li>Presented the platform to the department</li></ul>

Central NiFi Platform (HA Cluster)

Led and built a system that finds the same company across Finance's monthly SAP exports, puts it under its parent group, leaves out the company's own group, and adds up sales and purchases per group and currency.<ul><li>Works on any file layout: the AI sees only headings and digit-masked samples to find the columns</li><li>Seven steps: read, normalize, exclude, match, group, check, summarize</li><li>Every number comes from pandas. The AI only answers narrow batched questions and sees company names and countries only</li><li>AI answers are claims, checked against a YAML knowledge file that outranks them and open to a person's override. Stored answers make a second run take seconds</li><li>Every record shows how it was matched, who decided, and why</li><li>Web backend with finance, IT and admin roles: upload, run, review, release, and exports to Excel, PDF, Word and PowerPoint in four languages</li><li>Runs go through a NATS JetStream queue to Python workers, with Go services for the edge, AI gateway pacing, and worker scaling</li><li>Project manager: coordinated with Finance and designed the architecture</li></ul>

AIFinance: Company Matching for Finance

Built the backend for the Requirement page, where requests move through MES, AP, IT Server and OA review until the user verifies the result.<ul><li>Python FastAPI on an existing SQL Server table with SELECT and INSERT only: every change is a new snapshot row, and the current state is the newest row per request</li><li>Writes take an sp_getapplock lock and check the expected revision, so a stale edit gets a 409, and idempotency keys stop double submits</li><li>Workflow rules (allowed transitions, who may run each command, stage timestamps, task gates) live in a pure, tested module</li><li>Attachments in MinIO, with the preview type decided from the file's bytes and uploads bound to the request by HMAC-signed tokens</li><li>JWT identity, IT rights from the role table, and an encrypted connection string decrypted inside the container</li><li>Project manager: planned the work, coordinated with other departments, and supported users after launch</li></ul>

Requirement Request Tracking (Backend)

Built a shared Kafka platform that IT and other application teams plug their own apps into, instead of each team running its own broker.<ul><li>3-broker cluster in KRaft mode (no ZooKeeper), replication factor 3, every node running broker, controller, and JMX exporter</li><li>Separate tooling VM with AKHQ for topic and consumer lag inspection, Prometheus scraping broker JMX metrics, Grafana dashboards, and Redis for consumer services</li><li>nginx reverse proxy in front of all admin UIs</li><li>Specified broker sizing (4 vCPU, 8 GB RAM, 50 GB OS + 300 GB data, RHEL 8.10, non-root) and the full port and firewall plan</li></ul>

Central Kafka Platform

Debugging and protocol-tracing wrapper for stdio-based MCP servers, published on npm with 34 versions (currently 1.9.2).<ul><li>Relays server stdout to the client byte-for-byte while tracing every request and response with latency and protocol anomalies</li><li>Four subcommands: run, replay, stats (latency percentiles), and doctor</li><li>Fixed a request-id collision between client- and server-initiated requests, secret-redaction leaks, and UTF-8 corruption across chunk boundaries (Thai, CJK, emoji)</li><li>Fully automated releases with npm Trusted Publishing (OIDC, no stored token) and SHA-pinned GitHub Actions</li><li>Zero runtime dependencies</li></ul>