Some Past Projects
![](/assets/icons/mev-presentation.png)
MEV presentation - S21 Bali
Maximum Extractable Value (MEV) refers to the extra money that block producers can profit from by reordering or censoring transactions in addition to the standard block reward and gas fees. When a user sends a transaction to the blockchain, transctions sit in a staging area called a transaction pool (txpool or sometimes mempool) where their contents are visible to everyone. Searchers and miners monitor this txpool to find opportunities to make extra profits. We explore their various strategies both for sensing and execution.
![](/assets/icons/ontrac.png)
OnTrac
We won USD$6700 at ETHGlobal’s NFTHackathon for our Superfluid streaming-money accountability vault that helps you stay OnTrac with your most important tasks.
![](/assets/icons/manifold-optimization.png)
Introduction to Manifold Optimization and Geodesic Convexity
In this short report, we attempt to give a (very) brief overview of manifold optimization. Manifold optimization’s goal is to generalize optimization from flat Euclidean spaces to the larger domain of Riemannian manifolds. Not only does it promise to extend classical algorithms such as steepest descent and Newton’s method to a whole new set of problems, but also to rethink established solutions to classical problems, such as learning Gaussian mixture models, and in so doing outperform prevailing methods.
![](/assets/icons/reproducibility-hmc-nn.png)
Reproducing Generalizing Hamiltonian Monte Carlo with Neural Networks
As part of the ICLR 2018 Reproducibility Challenge, we reproduce the results from Learning to Hamiltonian Monte Carlo (Lévy et al., 2018), a general-purpose method to train Markov chain Monte Carlo (MCMC) kernels, parameterized by deep neural networks, that converge and mix quickly to their target distribution. The code for all experiments is available here.
![](/assets/icons/duckietown-slam.png)
Duckietown final project -- april-tags based SLAM
In this project, we wrote a SLAM algorithm for a duckiebot driving around duckietown using gtsam. Red arrows show the visual odometry, green arrows the optimized SLAM estimate, and green boxes the april-tags. After a loop-closure is detected, the SLAM algorithm gets optimized. In the video, we see the odometry drifting while the SLAM estimate corresponds well with the actual pose.