Well well. Welcome.
I am a domain expert in building and deploying machine learning models, vector search engines, and distributed computing/data systems. In 2021, I created a facial recognition algorithm that ranked US #1 and World #2 among top companies and research institutions, and have been maintaining it since then. In 2022, I drove the research and implementation of a new deca-billion-scale vector search engine, which has since been powering the service at Clearview AI. Being an academically trained computational physicist, I am passionate about performance computing, and the fundamental hardware/language/software breakthroughs that accelerate its progress.
I’m very hands-on with my work and believe in-depth experience and insights lead to strategic vision in this field. I enjoy the cutting-edge while seeking to understand established working patterns before instituting change. This technical Blog might interest you. Please find me on LinkedIn (CV).
EXPERIENCE
Clearview AI
Head of ML / Principal Engineer, 2021 - Present
I had previously worked on computer vision / facial recognition back in 2017, and really wanted to solve the two biggest problems in the industry - the accuracy of the algorithm, and the accuracy & scalability of the vector search engine at increasingly large sizes (deca-billion scale and more). Clearview was the perfect opportunity for me. I ended up solving them in each of the two years when I was here. In addition, I took on a lot more strategic and people responsibilities.
- Created a facial recognition algorithm ranked US #1 and World #2 (at time of submission) by NIST.
- Researched & implemented a deca-B vector search engine that was 100+ times the SOTA throughput. Fused high-profile open-source C++ libraries and created Python bindings. Near-term 1/10 server cost & much higher capacity. Long-term $multi-M annual savings.
- Achieved SOTA accuracy in the domain category with efficient model training orchestrating a cluster of nodes. Became early adopters of alternative chips, e.g. Intel Habana Gaudi2.
- Drove model efficiency using distillation training, quantization, pruning/trimming & hardware acceleration. Streamlined deployment with model encryption & serving engines.
- Established foundational ML/data practices & native (C++ & Rust) tooling for performance-sensitive tasks.
- Led cross-functional collaboration to design & implement multiple compute-intensive data pipelines & inference infra for images/videos, resulting in $100k cost reduction per recurring batch job.
- Co-wrote legal documents, marketing materials & public letters, engaged with policy officials to establish the company reputation.
MORE
During this time I got to be the sole person pushing through from data collection, pipelining, cleaning, preprocessing, training & experimentation, model productionization/inference optimization, all the way to indexing for search at deca-B scale. These tasks encompassed almost the entirety of the industrial ML/DL specializations.
My role here was a combination of ML Researcher (smaller part), ML/Data Engineer and Platform Engineer. The latter one echoed back with my experience at Bloomberg doing platform engineering work. In addition, due to the visual nature of computer vision, being able to anticipate and connect to the frontend/product experience was essential. Therefore I took lead roles at adjacent teams and contributed heavy-handedly to a few projects. The aim was to drive the initiatives to the logical conclusions and reap the benefits of what the ML core had to offer. Through them I got to really stretch my abilities. The hands-on technical experience and intuition I gained from these projects were priceless.
I carried various projects from research & prototyping to production in accelerated time frames, and obtained very concrete rules to determine progress in a sea of complexity. I made sure to communicate the expectations and stop gaps at each stage. With a few projects completed, I demonstrated that technology leadership could be driven out of a persistent vision.
Since the algorithm’s debut success, I became the trusted expert on the core tech in the leadership team and took on the scientific authority in product marketing and policy & legal messaging. I built the ML/Research division, laid out the technological foundation and assembled a team.
I encountered a vast lineup of technologies along the way, and became a proficient user in order to get the most out of them. They ranged from ML frameworks PyTorch, OpenCV, ONNX Runtime, OpenVINO, CuPy, TensorRT, approximate nearest neighbors libraries Faiss, hnswlib, and general SWE tools FastAPI, Socket.IO, Protobuf & gRPC, Redis Stream, RocksDB, Datadog, GitLab CI/CD, Docker, Kubernetes. I had to take face detectors and embedders apart in pieces and put them back together with different acceleration framework backends to understand how to create the best interfaces. I fused three high-profile C++ libraries (and more utility ones) together to arrive at what I needed for things to scale. Later I introduced a spectrum of Rust tooling into the stack for performance/cost-sensitive tasks. They were incredibly challenging and equally fun.
Bloomberg LP
Senior Software Engineer, Platform, 2017 - 2021
I moved into the industry and realized I really liked to work with computers in a more general capacity. In order to take things to the next level, I needed more rigorous training in Software Engineering. Here I became a more equipped developer, and learned on the job key data structures, system design patterns, essential tools and frameworks.
- Maintained & monitored widely-used bare-metal SFTP infrastructure with 7M daily logins.
- Designed & implemented reliable account management, auth, routing, caching, messaging for cloud-based next-gen SFTP. Wrote OS-level modules and web servers for auth interfacing with OpenSSH. Managed deployment on internal Kubernetes-based PaaS.
- Leveraged S3 Storage as SFTP subsystem with file system emulation and cross-data-center replication & failover.
- Spearheaded successful multi-year high-stake account migration as technical lead.
MORE
The work here felt like a high-stake crash course of the book System Design Interview. Clients, servers, DNS, API gateways, load balancers, proxies, route hosts, databases, caches, message buses, logging, metrics, rate limting, JWT auth, cryptographic padding/cert signing, user models, isolated networks and staged rollouts… and a lot more. It was humbling to see it first hand how seasoned engineers deal with large systems serving millions of users and requests every day. I made valuable contributions to the over-arching theme of transitioning from the old system to the new system, and matured into the likes of my peers.
At Bloomberg I became serious with C++ (it’s a C++ shop after all!), Golang, OpenSSH, Web development stack Flask, SQLAlchemy, Postgres, Redis, MongoDB, data/messaging frameworks RabbitMQ, Kafka, Flink, Spark, system monitoring tools Grafana, Splunk, Humio, DevOps & Platform Engineering tools Jenkins, Chef, Docker, Kubernetes, OpenStack. I took a special interest in scalable data/messaging pipeline and database components. Meanwhile I kept learning about ML/AI on the side.
University of Toledo
Ph.D., Physics, 2012 – 2017
Researcher & Developer, 2013 - 2017
I spent four of the five years here as a funded researcher working on First-Principles materials simulations with Density Functional Theory on supercomputing clusters. It was an exciting field that only became feasible after ~50 years of advances in supercomputing since WWII… not just quantum mechanics theories anymore. However the tooling around the core Fortran software package VASP was very scarce, and I had to be very proactive about what to learn and build. This journey taught me a lot of things, resilience being the most valuable aspect of it. Because of the fundamental similarity in problem scope (iterative optimization), working approach (scientific experiments) and technical requirements (big computers), it naturally paved my way to the growing Data Science and Machine Learning fields, and as soon as GPUs became more available, Deep Learning.
- Specialized in materials simulations with parallel computing clusters. 14 papers published.
- Calculated electronic ground states with gradient descent & residual minimization schemes, (non-)linear regression - similar routines as in ML/DL frameworks. Had various classical ML projects.
- Authored open-source projects: ScriptsForVASP, pydass_vasp, pyvasp-workflow.
- Built a materials database website with a modern stack, supporting tabulation/graphing, user auth/contribution. Project helped get a $100k research grant.
- Mentored three undergraduate and two graduate students.
MORE
Additional courses:
- Statistical Inference (MATH 8650)
- Categorical Data Analysis (MATH 8620)
- Topics In Statistics (MATH 8640)
- Statistical Consulting (MATH 8600)
Online courses:
- Introduction to Databases (Stanford) - with Distinction
- Algorithms, Part I (Princeton)
- Introduction to Probability and Data (Duke)
- Inferential Statistics (Duke)
- R Programming (Johns Hopkins)
- Regression Models (Johns Hopkins)
- Practical Machine Learning (Johns Hopkins)
- Statistical Learning “ISLR” (Stanford)
Dissertation on material properties prediction with cluster expansion regression models
Predicting properties from atomic configurations with linear models, cross validation and selection (cluster expansion formalism).
Here I became proficient with Linux and Python and their scientific computing eco-systems, e.g. NumPy, SciPy, scikit-learn, XGBoost, LightGBM, pandas, matplotlib, etc. I took a lot of additional courses and did projects to be fully equipped in SWE/ML/DL in the industry.
NOTE: if you would like to plan for a transition from PhD to the SWE/ML/AI industry, prepare early, and read A Survival Guide to a PhD from Andrej Karpathy and the HN discussion.
Nanjing University
Bachelor of Science, Materials Science & Engineering, 2008 – 2012
This major was interdisciplinary in nature. As a result, I studied a broad range of subjects: materials science & engineering, physics, chemistry, electrical engineering, computer science, math, statistics, etc. I also prepared and participated in a mathematical modeling contest with a team that focused on operations research.
MEDIA
(2023-06-28)
As explained in a company blog post by Liu and expanded on in conversation with Biometric Update, Clearview believes the smarter way is to index vectors so only small portion needs to be searched. This means “you can effectively search only a small portion of the database, finding very highly likely matches,” Liu says.
Blog Post: How We Store and Search 30 Billion Faces
(2023-04-18)
The Clearview AI platform has evolved significantly over the past few years, with our database growing from a few million face images to an astounding 30 billion today. Faces can be represented and compared as embedding vectors, but face vectors have unique properties that make it challenging to search at scale.
(2022-08-25)
Clearview Vice President of Research Terence Liu explained to Biometric Update in an interview that face biometrics algorithms are trained by ingesting several images from each subject, and then organizing data from ingested images into “clusters” with other images from the same subject.
(2022-08-10)
Clearview’s technology focuses on single images from commercial RGB images, VP of Research Terence Liu told Biometric Update during the same video call… Clearview takes an ensemble approach, combining models that look for different things, Liu says.
(2022-02-22)
Ton-That said Terence Liu is the Pennsylvania-based computational physicist behind some of Clearview’s algorithms and its vice president of research. They are listed together on a patent application filed Tuesday… The research head Liu formally joined last year after working as a senior software engineer at Bloomberg LP since 2017, he said. He earlier had partnered with Ton-That to advance Clearview’s prototype.
NYT: Clearview AI does well in another round of facial recognition accuracy tests.
(2021-11-23)
In results announced on Monday, Clearview, which is based in New York, placed among the top 10 out of nearly 100 facial recognition vendors in a federal test intended to reveal which tools are best at finding the right face while looking through photos of millions of people.
NYT: Clearview AI finally takes part in a federal accuracy test.
(2021-10-28)
In a field of over 300 algorithms from over 200 facial recognition vendors, Clearview ranked among the top 10 in terms of accuracy, alongside NTechLab of Russia, Sensetime of China and other more established outfits.