Aayam Shrestha

I build intelligent agents. 🧠💻🌳

Clean Code - Precise Writeup - Powerful Abstractions

2018 -
Ph.D. Candidate, Oregon State University
I am currently a Graduate Research Assistant at OSU advised by Dr. Alan Fern. My research aims to combine the reasoning abilities of symbolic AI planners / RL Agents with the rich representations learned by deep neural networks. This lets an AI Agent ground their abstract plans into the real sensory world - enabling optimal execution. I am interested applying my research to enable spatial computing agents to reason and act in virtual and real 3D environments.
SWE Intern (AI Specialist), Meta Reality Labs.
Designed and implemented Upselling module for Oculus Store. The Averagers recommendation module extends support for items, consummables as well as bundles for upselling. Offline evaluation demonstrated 150% improvement on suggested purchase conversion over KNN baseline. Also constructed User Journey Dataset for modelling in-app purchases over time across the oculus store.
Applied Scientist Intern, Amazon Search.
Designed and Implemented Page Level Reward for Cross-Slot Widget Ranking. This tackles the fundamental problem of defining reward which is aligned with different business metrics while factoring for page level interactions of users. Also built a prediction model that increases the modelling accuracy by 300% for rewards (over baselines) and 50% improvement for overall business metric predictions - across different world regions.
2018 - 2016
Business Intelligence Engineer , Logic Info.
Developed and maintained large-scale enterprise Data Warehousing and Business Intelligence solutions for Off-shore clients; Alex & Ani, Holland and Barett, Gander Mountain, and Makro. I was also invovled in Customer Experience Analytics over in-house data lake sourced from customer call audio, text reviews and curated twitter feeds.
2016 - 2012

B.S. in Computer Engineering, Tribhuvan University.
Discovered that a well written code teaches you how to think. Never looked back.


AMLC 2021 Workshop
Aayam Shrestha, Kai Yuan

ICLR 2021 (Spotlight Top-2%)   [Webpage]
Aayam Shrestha, Stefan Lee, Prasad Tadepalli, Alan Fern

Yodsawalai Chodpathumwan, Aayam Shrestha, Stephen Ramsey, Arash Termehchy
Pet Projects
DAC-MDP is able to solve Atari games using Tabular MDPs. It acheives this by building a Non-Parametric MDP over the learned deep representations. Ut then leverages GPU optimized VI solver from BIGMDP. It serves as a strong yet inexpensive baseline for small to medium offline Reinforcement learning Tasks.
BigMDP: A simple library for creating and solving large MDPs with million of states. easy to use APIs for MDP building and comes with GPU optimized VI solver. Able to solve MDPs with a millions of states in less than 30 seconds.
Mars Imageset classification using Semi Supervised classification. A large unlabelled dataset was first clustered using k-means and features from a discrete autoencoder. The clusters are then mapped to known classes using as small labelled dataset.
Obj2Obj GAN is a Masked Image to Image translation using GANs for object inpainting. This was primarily designed for dataset augmentation for generating images from different permutaitons of the object positions/attributes. Poster Image shows a horse inpainted with a zebra.
Bias in Knowledge Graphs attempts to quantify and generalize the difference in performance of machine learning algorithms over knowledge grapphs with information preserving structural variations.
Trained a Traffic Flow Prediction network using historical intersection data. The prediciton network consisted of a simple vanila RNN. The end goal was to optimize vehicle wait time using the learned prediction model.
Gurukul is a full fledge Enterprise Resource Planning platform with both mobile and desktop applications, built to be deeployt as SaaS. This was my Bachelors Final year project. It is capable of all normal administrative tasks, moreover it natively integrtates student data analytics to detect anamolies/weak students as well as provides data-driven recommendations for weak students.