I turn signal
into
insight.
About
I'm working on bringing responsible adoption of AI into our daily lives as CTO of GPTZero, the leading platform of detecting AI-written text. We help over 1 million users get transparency on the source of the content they read. If this mission resonates with you, we're hiring!Endlessly curious creator with a passion for deploying machine learning to solve impactful real-world issues.
My Interests: NLP and Misinformation // Autonomous Robotics // Synthetic Chemistry and ML
Experience
Enabling anyone to identify AI-generated text and certify human writing
- GPTZero is the world's most popular platform for detecting AI-generated text. We are a venture-backed startup whose mission is to foster an internet where authentic human content thrives.
- I lead our technical team in building state-of-the-art machine learning models, products that users love, and secure, scalable infrastructure.
- You can also find me working on our sales, customer success, ops, recruiting, and making sure the snack bar is stocked!
Building the next-generation autonomy for self-driving trucks
- I build Waabi's motion forecasting system to predict where other actors on the highway will be, so we can drive safely with them
- I develop and train highly efficient deep learning architectures that work at range, and use road priors to improve our system's accuracy and reliability
- I work closely with the motion planning and simulation teams to create end-to-end systems that optimize on-the-road performance, instead of isolated metrics.
Helping self-driving cars anticipate and avoid unsafe scenarios
- I developed deep graph neural network models in Pytorch to predict rare and unsafe behavior by pedestrians and drivers and plan safer driving for autonomous vehicles, reducing the number of potential collisions while more accurate modelling behavior from prior state-of-art models.
- To do this, I built new encoders and loss functions for sample-efficient, diverse, multimodal prediction.
- Our paper has been accepted to the 2021 International Conference on Computer Vision (ICCV) as an oral presentation (3% of all submissions).
Ads Growth Team attracts and assists new SMB advertisers
- I built a dynamic budget allocator for products that purchase millions of ads a day to optimize advertiser ROI.
- I added new data features and pipelines to improve our ads ranking neural net, increasing its conversion rate while increasing ad engagement.
- I discovered and fixed a major issue with a multimillion dollar ads ranking model. These neural nets rank Facebook notifications that aim to convert businesses to paying advertisers and reach millions of users per day.
- Finally, I created a social good project with 3 other engineers that won the Judge’s Choice award (given to top 3 teams) at Facebook’s largest-ever hackathon
Smart security cameras with machine learning-enabled detection
- At Kuna, I improved the speed, accuracy and cost of our machine learning cloud.
- I deployed an autoscaling, fault-tolerant AWS server scaler in production, decreasing company-wide cloud GPU costs by 60%.
- I sped up CNNs by 3x with quantization to achieve real-time detection on home security cameras.
- I built a statistically balanced dataset of 100,000 images and retrained models to reduce false and missed detections for users by 25%.
Web extension that generates a politically balanced, personalized news feeds, to improve media literacy.
- I built the ML + backend for a chrome extension with 100s of users, using Node.js and Flask for the backend, React for the frontend, and Spacy and NLTK for NLP.
- I presented our user studies on how to deal with polarization at Capitol Hill, Facebook's misinformation team, and the Department of State
- I and my team was interviewed by Fox News, AP, NPR and CTV.
- I managed and mentored a team of 3 developers to launch new features.
Project Page
Creating plug-and-play conversational and search AI for eCommerce
- I developed a chatbot platform for eCommerce for a Fortune 500 company, achieving 3x higher user engagement than the industry average with patent-pending NLP models.
- I built an ML pipeline to learn consumer product word embeddings and descriptors from open source datasets.
- I deployed an autoscaling ML computing cluster and continuous deployment pipeline in AWS.
Research
GoRela: Go Relative for Viewpoint-Invariant Motion Forecasting
ICRA 2023 (Award Finalist)
Alexander Cui, Sergio Casas,
Kelvin Wong, Simon Suo, Raquel Urtasun.
Efficient motion forecasting using shared, viewpoint-invariant scene encodings.
- In this paper, we propose an efficient shared encoding for all agents and the map when reasoning about the map, the agents' past trajectories and their interactions in order to produce accurate forecasts and enable safer driving.
- We leverage pair-wise relative positional encodings to represent geometric relationships between the agents and the map elements in a heterogeneous spatial graph. This parameterization allows us to be invariant to scene viewpoint, and save online computation by re-using map embeddings computed offline.
- We demonstrate the effectiveness of our approach on the urban Argoverse 2 benchmark as well as a novel highway dataset.
PDF Website
LookOut: Diverse Multi-Future Prediction and Planning for Self-Driving
ICCV 2021 (Oral, 3% selected from all submissions)
Alexander Cui*, Sergio Casas*,
Abbas Sadat*, Renjie Liao, Raquel Urtasun.
Contingency planning from diverse joint trajectory samples for all actors in the scene
- We train a deep graph neural network model to predict diverse predictions for the future trajectories of actors in traffic, such as rare and unsafe behavior by drivers, by rewarding it for predicting future scenarios that would force the self-driving vehicle to react differently.
- We pair this model with a contingency planner that can safely plan around unlikely, dangerous behavior by others, without overreacting.
- Our model reduces the number of potential collisions while more accurate modelling road behavior from prior state-of-art models.
PDF Video Demo
Multi-Label Classification Models for the Prediction of Cross-Coupling Reaction Conditions
Journal of Chemical Information and Modeling 2020
Michael R Maser*, Alexander Y Cui*, Serim Ryou*,
Travis J DeLano, Yisong Yue, Sarah Reisman.
Novel reaction-level graph attention model and data augmentation for learning reaction conditions
- We designed graph CNNs in Tensorflow to predict the best reagents for organic coupling reactions
- We optimized yield prediction of several reaction types with data augmentation and semi-supervised graph embeddings
Graph Neural Networks for the Prediction of Substrate-Specific Organic Reaction Conditions
ICML 2020 Workshop on Graph Representation Learning and Beyond
Serim Ryou*, Michael R Maser*, Alexander Y Cui*,
Travis J DeLano, Yisong Yue, Sarah Reisman.
A systematic investigation using GNNs to model organic chemical reactions.
- We compiled a dataset of four ubiquitous organic coupling reactions from the organic chemistry literature, with expert-clustered reaction conditions.
- We benchmarked 7 GNN models and identified specific graph features that affect reaction conditions and lead to accurate predictions.
Projects
Building autonomous submarines for the international Robosub competition
- Led ML team to train performant CNNs to locate objects with our autonomous submarine in C++, ROS
- Trained deep generative network (CycleGAN) to synthesize test-environment data to validate detectors in new underwater conditions
- Deployed adaptive thresholding, homography, and SIFT in OpenCV to track props with high precision
ML web app to generate photorealistic faces just from facial descriptors.
- Used generative adversarial network to reconstruct a face from only basic descriptions of facial features like age and hairline.
- Increased realism of generated faces with activation clipped
- Built the backend with Tensorflow, Flask, RabbitMQ
- Won Best Machine Learning Hack at Treehacks 2019.
ML web app to generate video cloning of faces, using any pair of face images and videos without further training.
- Combined First Order Model and facial sentiment classifier in Pytorch to generate video clones and detect user emotions.
- Built with Flask, Bootstrap, and Javascript.
I love hackathons! To see over 12 projects (including Best Hardware and ML Hack at Stanford's Treehacks 2021), check out my Devpost
Skills
Machine Learning
Pytorch
Tensorflow
Keras
Pandas
Full Stack
Node
Flask
AWS
React
Javascript
Bootstrap
Awards
Forbes | 2023
- Chosen as one of the 30 young leaders under 30 in Toronto, along with my co-founder Edward Tian, for our work at GPTZero
Citadel SoCal Data Open | October 2019
- Analysed factors that lead to Brexit, discovering the level of susceptibility to automation as being a key factor
- Competed against 25 teams of mostly graduate students teams across South California in a data science competition
- Our team was awarded $20,000
International Chemistry Olympiad | August 2015
- Selected to be one of the four people to represent Canada.
- Competed against the top chemistry students from around the world in theoretical and lab based exams.
- Mentored Team Canada in preparation for the 2016 international competition.
Education
Toronto, Canada
- M.Sc. in Computer Science in 2023
- Supervised by Raquel Urtasun and conducting research with Waabi Innovations
- 2021 Vector Institute Scholarship in AI
- 2022 Canadian Graduate Scholarship
Pasadena, CA
- B.S. in Computer Science, Minor in Data Science (3.9 GPA) in 2021
- Teaching Assistant for CS 155 Machine Learning and Data Mining and CS156b Caltech COVID-19 Prediction
- A Cappella, Ultimate Frisbee, Emergency Medical Responder, Student Waiter, Blacker and Avery House
Chicago, IL
- Study Abroad in Fall 2020
- Graduate courses in Machine Learning in Networks and Accelerated Computing
- Urban Design and Philosophy courses
- Political Union, Engineering Society, ML@UChicago, Eka House