Aditya Khandelwal

Software Engineer, TikTok

Mountain View, Jun.2020 - Present

•TikTok Business SDK for iOS
As the lead iOS engineer on this project, I am responsible for developing & maintaining the core measurement & ingestion logic required to send app event data to TikTok for attribution.

• Demographic Targeting of TikTok Users
I am responsible for building end-to-end inference pipelines that are used to produce gender and age labels for TikTok users. Currently, my work on these Machine Learning models is powering TikTok's Ads Engine to serve ads to millions of users every day.

Senior Capstone Project, Google & BMW

Stanford, Jan.2020 - Jun.2020

• Gesture Recognition for Vehicle Control
– Explored and prototyped a robust, production-ready, lightweight computer vision model for hand gesture classification
– Developed an end-to-end data collection pipeline, written in Python, to interact with Google’s Radar Kit
– Built an Android application in Kotlin to emulate a Heads-Up Display mounted in a BMW vehicle
– Implemented an MQTT broker to perform actions on Android application using hand gestures received through Radar Kit
– Interfaced with engineers and product managers at Google & BMW regularly to guide development cycle and use cases

Research Assistant, Stanford Internet Observatory

Stanford, Oct.2019 - Jun.2020

•Foreign Political Interference Detection Pipeline
– Utilized Apache Kafka Pub/Sub framework to ingest millions of Russian propaganda news websites to scrape. DevOps Tools used: Google Cloud Functions, Google Pub/Sub, Google BigQuery & Google Cloud Storage.
– Parallelized the scraping mechanism using a ThreadPool and load-balancing on Google Cloud Virtual Machines; reduced runtimes by 500% from several days to a matter of hours.
– Translated scraped documents, ran NLP techniques, like NER & Topic Modeling to present insights to social scientists.
– Created a web-UI to expose these research tools and datasets generated for larger audiences outside the research lab.

Sourcing Associate Intern, ArcForty Search Fund

San Francisco, Jun.2019 - Oct.2019

•Private Equity Data Pipelines
– Built a general-purpose, computer-vision based batch PDF scraper to spreadsheets tool accelerate last-mile reach-out.
– Provided insights on 700 potential American business and franchisee candidates that fit the search fund thesis.
– Managed a collection of Upworkers to inexpensively procure data on potential candidate companies.

Lead Software Engineer, Wanderwell

Palo Alto, Jan.2018 - Sept.2018

•Travel Recommendation Web & Mobile Platform
– Designed A/B tests for the initial iterations of the web & mobile products during the customer development phase.
– Proposed the design & implementation of a React-based Progressive Web App (PWA) to reduce development times.
– Increased Click-Through-Rates (CTC) on Facebook ads by 70% and Conversion Rates from DAUs to DPUs by 200%

Founder & CEO, The Great Indian Dream

New Delhi, Jan.2017 - Dec.2017 (Gap Year)

•College Admissions Platform For Indian Students
I was responsible for building a Kafka Pub/Sub pipeline to collect and process data from unstructured sources like foreign news websites and every relevant social media website for social scientists to perform qualitative analysis on. Key tasks included: creating general purpose scrapers, using Neural Machine Translation to translate text from hundreds of regional languages, creating a methodology for sentiment analysis on resulting data, building the infrastructure to automate and scale this process across websites and data sources.

Software Engineer Intern, LinkedIn

Jun.2016 - Sept.2016

•Internal Errors Visualizer
– Served as lead developer of a D3.js platform to visualize scrum code offenses, resulting in a 33% increased accountability
– Created a RESTful API to interact with this data and build more creative visualizations on top of it

Currently Working On: Using Generative Adversarial Networks for High-Quality Neural Style Transfer

Looking for research advisors & mentors

• Collecting high-quality artwork from museums around the world to train an Adversarial Network to produce high-quality artifically generated art segmented by regional art styles.
• Exploring metrics to differentiate GAN-generated artwork based on quality of images produced.

---

To Heat or Not To Heat: Reinforcement Learning for Optimal Residential Water Heater Control

Click here to view this paper

• This paper presents a reinforcement learning based approach for determining the optimal control strategy of a residential water heater. The objective is to minimize electricity costs while meeting temperature requirements when hot water is needed.

---

Simple Transformers for Personally Identifiable Information Identification

Click here to view this paper

• Data anonymization is a crucial prerequisite to clinical data sharing, transparency and follow up scientific analyses. Any such data shared must necessarily protect Personally Identifiable Information (PII).
• The goal of this research paper is to identify and extract named entities within discharge summaries by utilizing a novel NLP architecture, namely Transformer. Notably, I train a Transformer model on a large corpus of patient notes.
• This novel method is aimed at reducing the time taken to train the deep learning model whilst also achieving near-state-of-the-art results during prediction.

---

Atrous Spatial Pyramid Pooling for Semantic Segmentation

Click here to view this paper

• For this project, I present a deep learning framework that aims at solving semantic image segmentation of urban environments.
• To achieve this task, my proposed deep learning framework utilizes atrous spatial pyramid pooling. Training and evaluation is conducted on the Cityscapes dataset, which contains a diverse set of stereo video sequences recorded in street scenes from 50 different cities.
• Using the DeepLabv3 framework, numerous experiments on atrous convolution structures are conducted, and the final result receives a 72.53% mean IoU score on the Cityscapes set.

---

IdentifAI

Click here to view repo on GitHub

• This project focuses on providing an API for Locality Sensitive Hashing on "similar" images to assign ownership signatures

A decade ago, we partnered with Beachbody. From infomercial to altar, it completely changed our lives.

Today, we work with people who want to transform their health, make a living in fitness, or earn a little extra on the side.

Yes, it’s network marketing. No, it’s not gross. Because we don’t do gross shit. And we don’t team up with people who do.

Some say jack of all trades; we call it The Sickness™, a relentless compulsion to pursue everything to the max.

A DJ turned DEA Agent who flipped into fitness, I have a passion for writing and technology with roots in cyber security, web design, and brand development.

You might know me from various Beachbody workouts or my roll in developing the INSANITY LIVE format, but combine it all and you get an introspective health optimizer crushing excuses by conquering weakness.