The ERC-4626 Tokenized Vaults standard has helped to make yield-bearing tokens more composable across decentralized finance. The standard is optimized for atomic deposits and redemptions up to a limit. If the limit is reached, no new deposits or redemptions can be submitted. This limitation does not work well for any smart contract system with asynchronous actions or delays as a prerequisite for interfacing with the Vault (e.g. real-world asset protocols, undercollateralized lending protocols, cross-chain lending protocols, liquid staking tokens, or insurance safety modules). This standard expands the utility of ERC-4626 Vaults for asynchronous use cases. The existing Vault interface (deposit/withdraw/mint/redeem) is fully utilized to claim asynchronous Requests.

For the field of education, being able to generate semantically correct and educationally relevant multiple choice questions (MCQs) could have a large impact. While question generation itself is an active research topic, generating distractors (the incorrect multiple choice options) receives much less attention. A missed opportunity, since there is still a lot of room for improvement in this area. In this work, we train a GPT-2 language model to generate three distractors for each question, using the RACE dataset. Our model outperforms earlier work on distractor generation (DG) and achieves state-of-the-art performance. Next, we train a BERT language model to answer MCQs, and use this model to filter only corectly answered question. This improves not only our own results, but can also be used to enhance other question generation models.

For classical languages such as Latin and Ancient greek, determining morphology (gender, word type, etc) is of particular importance. However, most automated tools for this still rely on old machine learning techniques, rather than employing modern deep learning based models. We decided to create a BiLSTM based model, using the Perseus dataset for both languages. We trained a model which employed both word-level and character-level features, creating its own embeddings while training, outputting 9 different tags for each word. Our model greatly improves over previous works and achieves state-of-the-art results, with an overall accuracy over all tags of 96%, and a word type accuracy of 79%, for Ancient Greek.

We built a checkers playing robot, challenging people around the robot to play against it and taunting them when they were losing. We built a robot arm with Arduino components, which could move checker pieces using a magnet. The board was detected using a camera on top of the board, running computer vision to determine where all the pieces were. We used a minimax based game AI to determine the next moves, which was very hard but not impossible to beat. And while playing, a microphone was used to talk to the human opponent, giving you instructions, telling you when you made a bad move, and so on.