Aakash Mishra's Webpage
Recent Projects
Multi-Modal Langauge Models for Medical AI
We evaluate the performance of zero-shot classification models with domain-specific pre-training for detecting low-prevalence pathologies.
View
Distributed Traffic Control
We present a novel approach for mitigating traffic congestion using a distributed reinforcement learning model that is fault-tolerant and secure. Traditional methods of traffic signal timing optimization often rely on fixed timing schedules, which fail to adapt to real-time traffic conditions and can lead to increased congestion. In contrast, our proposed system leverages the power of distributed reinforcement learning to dynamically optimize traffic signal timings based on the current traffic patterns. The distributed nature of our model ensures fault tolerance and resilience in the face of individual component failures. By distributing the learning and decision-making process across multiple agents, the system can continue to operate effectively even if some agents become unresponsive or encounter failures.
View
RayICV (Ray ML Internal Cross-Validation Solution)
RayICV Implements a hyper-parameter tuning solution with cross-validation built-in on top of the Ray train framework in order to emulate the scheduling of jobs based on folds trained and model configuration.
View
The Data Open Citadel Boston Regional Datathon
Created a new variation of the statistical synthetic control method to apply K-Nearest Neighbors upon a dataset of 70,000 anonymized patient records from the NIH. We received this dataset while competing in the Citadel Correlation One Datathon and won $15,000 because we were able to show a causal effect in taking the FDA-approved drug known as Glipizide for diabetes and pulmonary heart failure.
View
Noether AutoDifferentiation Library
The NoetherAutoDiff Library is meant to be an efficient and approachable library for auto-differentiating differential equation matrices loaded as text / LaTex. Our library supports usage in both C++ and python (for Linux). The library is meant to be open-sourced and it can be installed as a module on PyPI (currently only for Linux).
View
Centaur Asteroid Research Project
We present multiband photometric observations of nine Centaurs. Five of the targets are known active Centaurs (167P/CINEOS, 174P/Echeclus, P/2008 CL94, P/2011 S1, and C/2012 Q1), and the other four are inactive Centaurs belonging to the redder of the two known color subpopulations (83982 Crantor, 121725 Aphidas, 250112 2002 KY14, and 281371 2008 FC76). We measure the optical colors of eight targets and carry out a search for cometary activity.
View
Robocup Harvard Robotics Project
Creation of robotic soccer players that emulate humans playing the same game. The goal of RoboCup combines a mix of automation, robotics, design, control systems, AI / ML and strategy. I lead the RoboCup team at Harvard and develop code for both the firmware and the high-level strategy code.
View
dftoltx
PyPi Library Module for taking Pandas tables and converting it to LaTeX formatted code.
View