I build backend systems that are supposed to be boring — the kind that hold up
millions of disks, millions of queries per second, and the
occasional very bad day. Today I work on Google Vids, an AI-native video editor.
10⁶+queries per second served by an ad-matching framework I launched at Google Ads
Millionsof disks managed by the Persistent Disk control plane I helped build
99.999%reliability for storage telemetry at petabyte scale, after an architectural
overhaul
8Mannual pageviews on a platform I directed with a team of 20+ engineers
Where I've shipped
Five years of building things that other people depend on — control planes, serving systems,
ML pipelines, and the product surfaces on top of them.
2026 — Now
New York, NY
Software Engineer — Google Vids
Google / Workspace Current
Google's AI-native video editor, part of Workspace. I work on the backend of its editing engine.
Audio arrives welded to its video clip. I designed the model change that lets it come apart — an audio clip whose backing data is still the video — and got it signed off.
The change wasn't local. The timeline, the export path and the clipboard had each been built on the old assumption, so all three had to move with it — inside an editor already in front of users.
BackendDistributed servicesAPI designAI product
Detaching audio from a video clip — and what it does to the model
2024 — 2026
Seattle, WA
Software Engineer — Storage Systems
Google Cloud / Persistent Disk
Google Cloud's block storage. Every disk on every VM is one of these.
Millions of disks, worldwide, under SLOs that assume the control plane never stops. I built the distributed services behind create, attach, snapshot, resize and delete.
Rebuilt the Storage Pool telemetry stack to five nines at petabyte scale — with a test story that did not exist before.
Customers would not sign until those numbers were trustworthy. That unblocked eight figures.
Wrote the authoritative design for block storage backup and restore — the system that matters the day something is lost.
And the unglamorous half: infrastructure config, productionization, and meta-monitoring — the monitoring that tells you your monitoring is lying.
Error budget per 30 days · what each nine costs you
2023
Sunnyvale, CA
Software Engineering Intern
Google / Cloud AI
Google Cloud's generative AI could not launch without customer-managed encryption keys.
I built the C++/SQL pipeline that backfilled CMEK across distributed databases. Zero data loss, zero user-visible disruption.
Owned it from first commit to launch. It cleared Agent Assist and Dialogflow.
C++SQLEncryption / CMEKOnline migration
2022
New York, NY
Software Engineering Intern
Google / Ads
A real-time bidding auction has milliseconds to decide which advert fits a slot.
I launched the C++ framework that matches ad-size requests to advertisers inside that path — millions of queries per second.
Ramped behind canary tests and global traffic-diversion experiments — in that path a bad match costs revenue on every impression it touches, so nothing goes out on a code review alone.
C++Low latencyRTBCanary / experiments
Slot size in, eligible creatives out — at auction speed
2021 — 2024
Los Angeles, CA
Director
BruinWalk / UCLA Student Media
UCLA's professor and apartment review site, run by UCLA Student Media.
8M pageviews a year. Part of UCLA Student Media's revenue depends on it staying up.
Director meant both jobs — roadmap, scope and stakeholders on one side, architecture and code review on the other.
20+ engineers through a full rebuild on Docker, Kubernetes and Nginx across DigitalOcean, AWS and Cloudflare. Active users up 10%.
KubernetesDockerNginxTech leadership
2021 — 2024
Los Angeles, CA
Undergraduate Researcher
UCLA / Medication Reconciliation Research Group
Two hospitals describe the same prescription two different ways. The gap is where patients get hurt.
I wrote the parser and reconciliation algorithm that reads natural-language prescriptions and finds the contradictions — a formal ANTLR grammar over medication instructions.
Shipped as a SaaS verifier on a UCLA TRAG grant, integrated with the US National Library of Medicine.
NLPPythonPipelines
2021
San Jose, CA
Software Engineering Intern
Cepton / LiDAR
A LiDAR unit is useless until you know exactly where it sits relative to the others.
I built the extrinsic calibration — machine learning over the point cloud to solve the transform between sensors. Beat the customer's precision spec by 60%.
Gaussian and kernelized models; K-means and bounding-box regression to clean the stream first.
PythonNumPyscikit-learnPoint clouds
Two sensors, one wall — solving for the extrinsic
Things I built because I wanted to
Side projects are where I get to pick the constraints. Lately most of them involve GPUs.
Generative ML
Photoreal diffusion pipeline
Trained a photorealistic personal LoRA on FLUX.2 Dev with ai-toolkit on an A100 80GB,
then built a JSON-driven batch inference harness for ComfyUI: one config declares the LoRA,
device, seed, steps, guidance, resolution and LoRA multiplier, with per-prompt overrides on
top. Hundreds of prompt/parameter combinations render unattended overnight instead of
one-at-a-time in a UI.
Quantum computing
Shor's algorithm, three ways
Order-finding implemented in Qiskit three times over — the textbook circuit, a
phase-estimation version with the redundant unitary stripped out, and an iterative version
that reuses a single ancilla qubit — then run noiseless and again against
FakeMelbourne, which models the real error rates of IBM's 16-qubit machine. The
textbook circuit's noisy output is unusable static; the trimmed circuit gives clean peaks;
the iterative one needs so few qubits that both runs finish in ten minutes. Factored 15, 21
and 63. The honest conclusion was that the algorithm is fine and the hardware is not.
A file-transfer client and server over raw UDP in C++, with no help from TCP: selective repeat
with a per-packet timer, receiver-side reordering by sequence number, and AIMD congestion
control — additive increase of one packet per RTT, halve the window when a packet outlives its
TTL. Tested against a link simulator that drops, delays and reorders traffic on purpose —
written for a networking course, and still the clearest thing I've built about why TCP is
shaped the way it is.
Parallel computing
One convolution, four machines
The same dense linear algebra across four execution models, each with a different notion of
what "fast" costs. OpenMP: 0.62 → 115 GFLOP/s on a 1024³ GEMM through loop
permutation and cache blocking. MPI: Cannon's algorithm scaling near-linearly to 88
GFLOP/s at 4096³. CUDA: a CNN kernel from 51 → 585 GFLOP/s, most of it from
keeping the accumulator in a register. FPGA: the same kernel through Xilinx HLS onto
a U200 at 243 MHz, where the currency is area rather than clock. The loop-order half of
that is taken apart in the lab.
B.S. Computer Science — UCLAGPA 3.99 / 4.00 · Summa Cum Laude · Dean's Honors · 2020–2024
USA Computing Olympiad — Gold DivisionTop competitor nationally
Ann Tsu-Ann Tai & Kam Sing Tso Centennial ScholarshipAwarded in Computer Science, UCLA
4th Place Team — Texas UIL InvitationalComputer Science, Houston, TX
CourseworkAlgorithms · Operating Systems · Databases · Machine Learning · AI · NLP
Contact
Let's build something hard.
I'm most interested in high-leverage backend work — inference and training infrastructure,
distributed storage, and low-latency systems where correctness and speed both actually matter.
If that's the kind of problem on your desk, send me a note and I'll reply from my own address.