Kunanon
Thappawong

Focus

Research interests

Multimodal AI that understands people, and AI that leaves the lab: running on small devices, in homes, factories, vehicles and robots, and supporting health and wellbeing.

Education

Education

• Apr 2023 – Apr 2027
  (expected)
  

  B.Eng. in Computer Engineering

  Mahidol University International College (MUIC), Thailand · Track: Intelligent System and Robotic Sciences

  • 50% tuition scholarship for all four years, Computer Engineering International Program.
  • Languages: Thai (native), English (advanced).

  Selected coursework

  • Embedded Systems
  • Microprocessors and Interfacing
  • Computer Architecture
  • Mechatronics
  • Operating Systems
  • Computer Networks and Information Security
  • Database Systems
  • Linear Algebra
  • Artificial Intelligence
  • Human Robot Interaction
  • Data Structures and Algorithms

  Emblem: Mahidol University, public domain, via Wikimedia Commons.

Research

Research experience

• May 2026 – Jul 2026

  Research Intern, KnOWLab

  Japan Advanced Institute of Science and Technology (JAIST) · Project “Real-Time Emotion Analysis in Video Calls” · Supervisor: Assoc. Prof. Natthawut Kertkeidkachorn

  • Designed a controlled benchmark of Qwen3-VL-8B (chosen over LLaVA-NeXT and LLaVA-OneVision in screening) against EmoNet and an MLP trained on extracted features, using unscripted Thai video-call clips, for a task with four circumplex emotion classes.
  • Built a multimodal fusion pipeline that places five extracted cues (facial action units from Py-Feat, Whisper transcripts, OpenSMILE voice acoustics, gaze and head pose) in the prompt together with the video, and ran every combination of cues under two prompting strategies on the JAIST GPU HPC cluster.
  • Measured the marginal contribution of each cue across all combinations, rescored every system configuration from raw predictions on one shared clip set for a fair comparison, and built interactive HTML and JavaScript dashboards to compare runs.

  

Projects

Selected projects

• Aug 2026 – present

  Alani: Local Voice Assistant and Discord Automation Platform

  Personal project · Python and Node.js · more than 9,000 lines of code, more than 150 commits

  • Voice assistant running fully on a local RTX 3060 Ti with 8 GB of VRAM: a custom “Hey Alani” wake word (trained on about 20,000 synthetic clips), faster-whisper speech recognition, Qwen3-8B through Ollama with tool calling, and Chatterbox Turbo speech synthesis with zero-shot voice cloning (4 voices, faster than real time). An idle listener costs almost nothing and starts the heavy pipeline only when needed.
  • Designed 13 tools for the language model (reminders, calendar events, web search, weather and more) and a policy of refusing unsupported requests instead of guessing. Diagnosed and fixed a runaway loop caused by speech recognition hallucinating on silence.
  • Built an Electron and Three.js status widget connected by WebSocket, with settings, a typed input mode and manual reminder and event forms.
  • Discord bot (Node.js), hosted around the clock: a three-tier permission database with collaborators, reminders and events synced to calendars and Discord scheduled events, a music player, health and routine trackers, and scheduled daily jobs.
  • Deployed the JAIST research pipeline as a web service with two-step valence and arousal prompting, Py-Feat, Whisper and OpenSMILE: an HTTPS dashboard for uploading or recording a video, with live progress and a results page (Alani Emotion).
• 2025

  University Assistant: Local RAG Chatbot

  Artificial Intelligence final project (team of 4, primary developer)

  • Built a question answering system over university documents using retrieval augmented generation, running entirely on local models: text extraction and OCR (pdfplumber, Tesseract), chunks of 500 tokens, MiniLM embeddings, a FAISS index, and 4-bit Mistral-7B-Instruct (llama.cpp) on CPU.
  • Moved the pipeline from the OpenAI API to open-source models after API quota failures, removing the paid dependency and keeping documents on the local machine.
• Mar 2025

  Smart Mask Box and Line Following Robot (Raspberry Pi)

  Embedded Systems final project

  • Python mask dispensing box that monitors air quality (Sensirion SPS30 particulate sensor), estimates mask stock with an ultrasonic sensor, logs door events with infrared sensors, and reports status on an LCD and LEDs with a button-driven settings menu.
  • Motor control for a line-following maze robot: H-bridge PWM speed control, steering from three infrared line sensors.
• 2025

  Machine Learning Coursework

  Artificial Intelligence course

  • Airline passenger satisfaction classifier (decision tree tuned with 5-fold grid search), a Quick Draw sketch classifier (Keras CNN) and a hill-climbing agent for 4 by 4 tic-tac-toe.

  Also: a parcel delivery simulation with Java concurrency, Maze Runner Quest (a Java Swing game), SQL and MongoDB aggregation work, a chess game in C, and game design and digital art with Inkscape and GIMP.

Skills

Technical skills

Beyond the code

Leadership

• Vice President (2 years)
  Secretary (6 months)

  MUIC AGC Club (Anime, Games and Comics)

  • Worked with 30 other student staff to run events on anime, games and comics for MUIC students, including many in-person and seasonal events.
  • Sole organiser of a competition event with 28 participants and a 5,000 baht prize pool.
  • Ran a Minecraft competitive event series over five seasons that drew 897 total online views within the university.