I'm currently an incoming CMU
MSML student researching under Jun-Yan Zhu in the Generative Intelligence
Lab.
I'm also a head TA for 15-151 Concepts of
Mathematics.
Here's a brief overview of my work. My Resume again,
for reference.
Education
Sept 2023 - May 2024
Masters of Science in Machine Learning Carnegie Mellon University
My senior year I took Cognitive
Robotics, a course in which you program Cozmo, a robot with a
camera sensor.
Thanks to CMU, cozmo also had access to a ~8 GB GPU. For our final project, my partner
Akshath and I decided to use MiDaS,
a monocular depth model,
to predict depth at every frame that Cozmo sees.
Since MiDaS only gives relative depth,
this depth map is not grounded with real world depth values. However, when Cozmo sees a
light cube, a special object with an aruco marker,
he knows how far away this light cube is. Using light cubes as a sparse depth signal, we
calculate an optimal scaling factor to multiply to the relative MiDaS depth map to give an
accurate depth map of the image, which
can then be queried at any pixel. Feel free to look at the slides linked above for a full
explanation and proof of optimality for our scaling
factor!
Battlecode is an MIT AI competition run every year throughout the month of January with
100s of teams and thousands of participants entering code. As a general overview, games consist
of two teams, each with control of some number of robots. These robots have different
abilities(making money, attacking, creating more robotos, etc),
and can only communicate through bitFlags whose messages must be coded up in some finite range. Some strategies for performing well include implementing fast pathfinding to navigate terrain with different levels of movement allowed per square, clustering troops to place them strategically across the map, and using map symmetry to infer the location of the enemy base long before actually exploring the entire map.
There are multiple tournaments throughout the month, ending in a tournament for the top 16 teams
getting flown out to MIT for cash prizes. We were lucky enough to be
in the top 16 for the past two years, with monotonically decreasing rankings for each year. Our first year, we placed 9th-12th, the best performance out of all first time teams. The next year, we placed 7th-8th, and our final year we placed 3rd.
This year, we worked really hard to make a post mortem that is both easy to read and
informative for any level of reader. Please feel free to give it a skim/read some paragraphs
that are of most interest (the table of contents has links to each section, so you can just
click on a section to go there)!
Here are some of the cooler features from this year:
Like in 2021, we used a stack to store states (look at the first bullet 2021 for more
info)
We used a distributed k clustering algorithm for troop movements. All troops reported
enemy troops in their range to the main tower, which then found at most 3 main enemy
clusters. From here our troops went towards their closest cluster
We spent a lot of time on troop to troop micro-interactions. for deciding whether to
attack at a micro level, we considered how many troops we had vs how many troops the
enemy had as well as health, cooldown, and land passability
Here are some of the cooler features from this year:
Our home base towers used a stack to store different States. With this in place, we
could switch from a state like Default to defending, while pushing Default to the stack.
Once we were done defending, we could pop the State Stack and go back to Default mode.
This allowed us to do a lot of tasks as intermediates while still having main tasks
We used priority queues that stored locations of enemy bases, giving priority to those
that had the least amount of money, to effectively take over enemy bases when possible
We used bit manipulation to communicate between towers, which all had 24 bit flags that
robots in sensor radius could see
We used the Spotify API in conjuction with Google's NLP API to create an app that
classified the mood of your songs. In addition to classification, we were able to query certain
moods from
a user and output songs that had the most correlation to the mood that the user wanted.
Implementation
Given a user mood, we first transformed it into a mood vector.
From there, we used Spotify's API to find a subset of songs with similar mood vectors. From
here, we created a further embedding that used both spotify's
mood vector and our sentiment analysis result and found the song with the highest dot product
with the original user mood vector
We also used a similar process to generate a curated playlist given a specific mood, as well as
added ability to log in to your current spotify account and save said playlilst.
Finally, we created a function to plot mood over time, so the user could see how the mood of
their music changed over some time period.
The main goal of this project was to take data from the regular season of the NFL and
predict the playoffs be computing rankings for each team.
Implementation
We employed two main strategies:
In the first, we try to find rankings of teams, such
that for every game between teams i and j, the difference between team
i's score and team j's score is equal to their difference in ranking. This method didn't work
well at first, but after taking all nonzero terms in our matrix and making them small values
10^(-15) it was very efficient. We surmise that this is because we were able to increase the
stability of the matrix through perturbation
In the next method, we created a matrix that is 32 by 32, such that every entry Aij is equal to
the total number of points scored by team i against team j,
with some normalization. Next, we tried to find some strength vector S such that,
when multiplying A by S, we get a vector proportional to S. This becomes an equation where we
can use eigenvalues and eigenvectors to find a solution. It
worked better than a naive version of strategy 1, but worse than the perturbed values.
Our goal was to create a program that, given a start and end location, gives a user the
safest timely walking path to their destination. We used Manhatten as a test area since they had
a lot of crime data for us to use.
Implementation
First, we downloaded all street nodes from OSM, a mapping API. Next, we assigned each node a
value based on the amount of crime in the area, specifically weighting crime intensity via
a function that took in severity of a crime as well as decayed in value deending on distance
from a node. We also only considered crimes within a certain latitude longitude value of a given
node.
From here, given a path, we would then run an A* search on the start and ending nodes, with the
weight of each node being the crime rating we had assigned.
All nodes used for our A* Search
Heatmap of Manhatten, with red/brighter areas representing more
crime
The idea of this project was to create a platform that origamists could share their work
with each other and also share models that they made with each other along with
instructions
Implementation
I first added a login feature with hashing passwords using MD5 so users can be authenticated in
a secure manner in the backend. From here, users can look at the designs of their freinds,
add their own designs, and create new designs via a python program embedded into the webpage.
Origami
Models
I've been folding origami for over a decade now, and over time
I've been able to become skilled enough to fold some great models.
Nowadays I do most of my folding at CMU's origami club. Here are some of my favorites from my
origami instagram
page
Teaching
I've been teaching origami since I was in middle school. Here are some pics of me teaching
from when I was younger:
Me in 8th grade teaching origami to kids in South Carolina
Me in 9th grade teaching origami to kids in Kenya
Me in 10th grade teaching origami to kids in Myanmar
Basketball
I've been playing basketball since around second grade, and I currently play club
basketball at CMU. One of my big goals for basketball was being able to dunk. Here's a video of
me dunking :)