Naphat Tangpanitchakorn

My largest personal project is a Thai-speaking AI VTuber inspired by Neuro-sama. Karu can listen to my voice and read live YouTube chat, generate responses based on her own personality, and reply using a cloned voice with a Live2D model. She can co-host streams with me or run them independently. Her Thai is still improving, and I continue developing her over time.

• ▸Real-time pipeline: speech-to-text (faster-whisper), a persona LLM, Thai voice-clone TTS (F5-TTS) with emotion tags, and Live2D lip-sync, all streamed through OBS
• ▸Memory system: embedding retrieval, pinned facts, and daily self-generated 'thoughts'
• ▸Fine-tuned a local LLM on a dialogue dataset I built by hand (QLoRA to GGUF, served with Ollama)

A comprehensive laboratory management platform for managing users, computer labs, devices, class schedules, and reservations. It integrates with a Windows-based Tracking Agent for user authentication, real-time device monitoring, and centralized computer control. The system is designed for organizational use and scalable deployment across many computers.

• ▸Reservation and walk-in management with availability checks, class schedule conflict detection, duplicate booking prevention, automatic cancellation, and penalties for reservation no-shows.
• ▸The Windows Tracking Agent requires users to log in before use and enables remote computer control from the website.
• ▸The centralized management system covers users, role-based permissions, rooms, labs, devices, class schedules, notifications, and device blocking.
• ▸Face recognition identity verification system that detects and verifies users when entering and leaving the laboratory, with anti-spoofing protection against fake images.
• ▸AI-powered website monitoring and classification system that analyzes accessed websites in real time, automatically categorizes content, and immediately alerts administrators when potential risks are detected.

A storefront I built for a wooden furniture shop in Ayutthaya. Customers browse the catalog and pick what they like, and the buy button opens a chat with the shop's page. That's how this shop actually closes sales, so I built the site around their real workflow instead of forcing online payment into it.

• ▸Product catalog with categories, SKU search, and a ready-to-ship filter
• ▸Admin back office for adding products and promotions, with an overview dashboard

A platform that runs in production with real paying customers. People register, top up with PromptPay, and buy usage hours for a desktop program I made. Running it taught me things a tutorial never could: when something breaks, a real customer messages me about it.

• ▸PromptPay QR top-up with hour packages, custom amounts, and full history
• ▸Admin console for users, staff, pricing, and online-user stats
• ▸Push program updates and kill-switch clients remotely from the web

An IoT prototype that watches for falls. An ESP32 reads acceleration from an MPU6050 and looks for the pattern of a fall: a hard impact followed by the person lying still. It warns on the OLED, LED and buzzer first, and only sends the alert if nobody cancels within ten seconds — so a stumble doesn't call for help by mistake.

• ▸Rule-based detection on total acceleration (impact over 2.5g, then stillness), driven by a six-state machine from NORMAL to ALERT SENT
• ▸Cancel window before the alert goes out, to keep false alarms from getting through
• ▸Web dashboard over the ESP32's HTTP API: live acceleration graph, adjustable thresholds, and scenario buttons for demos
• ▸Runs end to end in the Wokwi simulator via PlatformIO, so the whole system can be demoed without physical hardware

A university project (CS460) that teaches at the level you're actually at. You pick a subject — or add your own — attach your course files, and ask a question. Gemini reads your answers to work out where you stand, then the app explains the topic at that level and hands you practice questions to match.

• ▸Assess → explain → practice flow, with Gemini analyzing the answers and setting the level for the explanation that follows
• ▸Upload course material (PDF/TXT/MD) as context, and add custom subjects saved on the device

A hidden-role party game for one phone passed around the table. Before each match the app asks OpenAI to write the questions and draw an illustration for them, matched to the category and the age group you pick (Kids, Teen, Adult). One player secretly becomes the Hunter, one the Blue Fish, and the rest are Red Fish — the story phase, the hunt, and the scoring all run inside the app.

• ▸AI-generated content per match: questions written by gpt-4o-mini, then reviewed and illustrated with a generated image
• ▸Rule engine that assigns hidden roles and settles the score for every round
• ▸Match history, session detail, and a leaderboard kept on-device in SQLite

A Roblox map where players summon an NPC to dig for treasure on their own plot, with a gacha roll on what comes out of the ground.

A Roblox map built around a summoning mechanic: players call random stones up out of the ground.

A top-up and shop system running inside a FiveM game server. Players scan a PromptPay QR to pay, then submit their slip with the date and time from the in-game UI. The request goes straight to the admin team's Discord for approval, and once approved the points land in the player's account to spend in the shop.

• ▸Slip submission from the in-game UI that notifies admins on Discord for manual approval
• ▸Point wallet and item shop — players spend approved points on in-game items
• ▸Live status panel showing server state, point balance, and total donated

A single hub that gathers every feature on a FiveM server into one screen. Players used to memorize a different command for each system, so I put them all behind one menu with a short Thai description under every button — including a status label for the features that aren't open yet.

• ▸12 systems in one grid — store, daily login, redeem code, battle pass, party, airdrop, camera, and more
• ▸Every button carries a one-line description, so new players can find things without asking

A 28-day login reward system for a FiveM server. Players log in each day to claim that day's item, and the calendar shows at a glance what's already claimed, what's available today, and what's still ahead. Claim history is kept per player in the database so the streak survives restarts.

• ▸28-day calendar with three clear states: claimed, available today, and still locked
• ▸Per-player claim records in the database, so progress carries across sessions

A desktop app for managing stock in a small shop or warehouse. Staff log in with their own roles, add products, and move stock in and out. Every change gets posted to the team's LINE chat right away, so the owner always knows what happened without opening the app.

• ▸Product and category management with SKU search
• ▸Stock-in / stock-out records for every movement
• ▸Admin and staff roles with full user management
• ▸LINE notification on every stock change

A bigger desktop inventory system that covers the whole warehouse side: suppliers, customers, products, storage locations, and staff, with inbound and outbound stock tracking on a local SQLite database.

• ▸CRUD screens for suppliers, customers, employees, products, and storage locations
• ▸Inbound / outbound records with quantity and unit price per line item
• ▸Activity log that keeps a timestamped record of every add, edit, and delete
• ▸Themed desktop UI (ttkbootstrap) in a master-detail layout