Portfolios
Nomad dataset challenge
Transparent conductors are an important class of compounds used for photovoltaic cells, light-emitting diodes for flat-panel displays, transistors, sensors, touch screens, and lasers. However, only a small number of compounds are currently known to display both transparency and conductivity suitable enough to be used as transparent conducting materials.
This project works as a quick screening for potential for finding new materials with an ideal target functionality.
Named Entity Recognition and Disambiguation (NERD) with Wikidata
This is my excuse to learn ReactJS, Docker, and deploying with EC2. I played around with online demos of NERD and found the performance a bit unexpected.
Is it possible to do NERD in short queries, where there are fewer grammar clues or capitalization like in a sentence?
Would watch [harry potter] vs read [harry potter] be resolved to Harry Potter film and Harry Potter book?
I create a demo. I discovered that there are fewer datasets that deals with short query than I thought.
Hardware: 2 GBs RAM, 1 shared CPU
What I like: I devise tests to see if the cosine distances of different Word Embeddings suit the short text nature.
[Code]
Master thesis
Gas price prediction
What I like: I stepped away from accepting the model as the black box and tried to wrestle some sense out of it. Below are the ten most influential words to gas price according to my model. It is Table 6 in my paper.
| 2012 | 2018 | |
|---|---|---|
| 1 | oil | energy |
| 2 | energy | gas |
| 3 | price | oil |
| 4 | FTSE | China |
| 5 | fall | Trump |
| 6 | shale | trade |
| 7 | power | price |
| 8 | coal | LNG |
| 9 | deal | UK |
| 10 | Shell | rise |
[Paper with code] [In Proceeding of CEUR Worshop]
Rapid road detection
Road detection with 50% faster interpretation speed and retain more than 95% quality.
What I like: My first computer vision project. Seeing the result of the work in the visual form was quite a sensation.
Industrial robotics software
Pose estimation on a tabletop environment challenge, whose whole dataset consists of two images. My work helped the University of Bonn’s win the EuroC Challenge.
What I like: How I adapted when the training data consists of RGBD data of one image and still deliver (Citroën bought our work!)
Health hackathon
In 48 hours, my team created E-Prescription, an Android app to remind patients to take their medicine.
What I like: In Vietnam, antibiotics are overused as one does not need a doctor’s prescription to buy some. When they forgot to take their medicine, it is common to double the dosage in the next intake. It contributed to Vietnam being among the most overly tolerant antibiotic countries.
[Our app featured in Tech in Asia][Facebook page of the Hackathon]