AnthonyJiaqi Chen

Reputation Passport

Portable seller reputation / Independent

Small ecommerce sellers build reputation on one marketplace and lose it the moment they expand to another. Sellers connect their storefronts, Shopify first; the service imports and verifies sales, ratings, reviews, and dispute history into one portable profile that marketplaces query through a paid REST API. Every imported record and every score is Ed25519-signed and appended to a hash-chained attestation log, so a passport verifies offline against a published public key. Tamper-evident without a blockchain. Go API with River workers, Next.js dashboard and badge embed, Postgres 17.

• Go
• Next.js
• Postgres
• Ed25519
• REST API

sourcedemo

sha256 21ff6c126e98fe1783f3c5865743182267cb05f20f31683bb0696f35e5f00112

Site Reliability Engineering Intern, Data Infrastructure

ByteDance / Bellevue, Washington

The Byte Message Queue team’s cluster management platform broke during a metadata database migration, and the 30 engineers who depended on it had no documentation for it. I traced live network API calls to recover what it did and rebuilt that functionality against the new backend, then designed a single interface for upgrading clusters, scaling cluster counts, and reaching Grafana — replacing a redundant legacy workflow. Cluster health had been computed on demand; an hourly cron now precomputes disaster-recovery data into a metadata layer and the front end surfaces scalability-guideline violations directly.

• 2 min → 200 ms cluster health
• 53% faster instance statistics
• 30-engineer team unblocked

• Go
• message queues
• Grafana
• observability

sha256 2ef317d08c2268aa34c2c2e8f9e8bbf5382c836ffc646d3db63f5c8d51337325

Village on Wimpy Nodes

Prescience Lab, Northwestern University / Evanston, Illinois

Village is a data-parallel language compiler and runtime built on NESL. The question was whether it can run on hardware nobody would call a supercomputer. I built the cross-compilation workflow from x86 to aarch64 — extracting fgraph binaries and applying them to a precompiled runtime — which got a 20-node Raspberry Pi 5 cluster, 16 GB each, executing at 98% RAM capacity. The dynamic partitioning scheduler I wrote runs 6.7 times the efficiency of round-robin FCFS by voting on data locality to minimize cross-node transfer.

• 6.7x scheduler efficiency
• 98% RAM utilization
• 20 nodes over MPI

• MPI
• compilers
• aarch64
• distributed scheduling

source private

sha256 eca4154b8acb50fd2ad43877bd864af3d76e12861adff90df173b0e52c45af34

StreamSRE

Data-infrastructure reliability lab / Independent

A self-contained streaming stack built to practice the failure modes rather than read about them: Go producer into Redpanda, Go consumer into Postgres, with Prometheus and Grafana on top. Idempotent consumption through a processed-events table, dead-letter routing for poison messages, chaos scripts, and incident runbooks for consumer lag, slow database retries, and DLQ pileup.

• Go
• Redpanda
• Postgres
• Prometheus
• Grafana

source

sha256 2d4ca33b05c3ba789dd02ed28ec26be171926aee86373eaba4a4223e50eb4d25

Research Intern, Cybersecurity Lab

Swinburne University of Technology / Melbourne, Australia

Replicated and verified two deepfake-detection experiments on the university HPC network, then stress-tested their reliability against a custom 1,000-photo dataset in PyTorch and documented where detection fails to transfer between GAN and Diffusion architectures. Built the Python framework that automates preprocessing and metric logging so the benchmarks reproduce.

• 3 models benchmarked
• 1,000-image custom dataset
• 40% less setup time

• PyTorch
• HPC
• reproducibility

sha256 90358a12a911726bf77fad33fc46137a74ede1c8f79ed3f7ef7e600b349a5444

FicRead

AI narration platform for web fiction / Independent

The predecessor to Bookhush, and where the audio pipeline was worked out. React and Material-UI front end, a FastAPI narration service driving ElevenLabs voice synthesis with per-passage emotional tuning, and AWS Cognito, Amplify GraphQL, and DynamoDB behind it. Multi-stage Docker builds cut the deployed image by 93% before it shipped to EC2.

• 93% smaller container image

• React
• FastAPI
• ElevenLabs
• AWS Cognito
• Docker
• EC2

source private

sha256 57d5b04e321699901ccd04e45f9b2268da93d475bd13543183dfd42ab853f131

NU Course Recommender

Course discovery for distribution requirements / Northwestern University

Northwestern students have to fill distribution requirements and have no good way to find ones they will actually enjoy. A scraper keeps course data current in DynamoDB; a K-nearest-neighbors similarity model in PyTorch ranks courses against a student’s stated interests, served through Lambda behind a React front end.

• Python
• PyTorch
• KNN
• AWS Lambda
• DynamoDB
• React

nu-distrorecs.comsource

sha256 75958148db1d7422003e3f889fe787cd49086aa281b800e9ea6957e9ea13a5dc

Bookhush

Web novel reading platform / Independent

A full-stack reading platform with real-time text-to-speech over more than 50 GB of audio and text. Vite and TypeScript on AWS Amplify — GraphQL API, Cognito auth, S3, DynamoDB — in front of a Dockerized microservice architecture whose Python FastAPI service runs the Piper neural TTS engine.

• 95% cache hit rate
• sub-200ms TTS responses
• 50GB+ served

• TypeScript
• FastAPI
• AWS Amplify
• Docker
• Piper TTS

bookhush.com

sha256 7fb10236a063b887c2c7e4e75c4b0a49d291aa9ced5b950a4f38a0e43a0f8768

Northwestern University

BS/MS Computer Science / Evanston, Illinois

Combined bachelor’s and master’s in Computer Science. Peer mentor for CS 150. Master’s thesis on stability in dynamic coalition formation of altruistic hedonic games, advised by Edith Elkind.

• GPA 3.87 / 4.0
• Dean’s List, 5 of 6 quarters

Freshacks Hackathon, first place

• algorithmic game theory

sha256 bf7b961e1fc21059500463f97df3b498d08df83c7665d68e61e6ba245ecc9ec2

Software Engineer Intern, Kernel and Platform

Oppo / Shanghai, China

Adapted two open-source scheduler benchmarks for the Linux kernel; the resulting internal tool is now used by three teams and more than sixty engineers. Tuned process threads, mutexes, and semaphores — including an optimistic spin lock implementation — for a 10% throughput gain under multi-core workloads on a new handset. Wrote the 15-page technical guide and the Bash automation for calibration, test runs, and result logging.

• +10% multi-core throughput
• 60+ internal users
• 20 hrs/month of manual testing removed

• Linux kernel
• C++
• Bash
• concurrency

sha256 4e35a3370de7bd034ce05cbfadc67345e82bf84d3c4338af4209e37e16835a1c

University of Michigan, Ann Arbor

BSE Engineering / Ann Arbor, Michigan

Two semesters of core engineering before transferring. Graduated the program with a 3.96 GPA as part of MRADS, the Michigan Research and Discovery Scholars.

• GPA 3.96 / 4.0

William J. Branstrom Freshman Prize

sha256 4f125622bd72ac095b1dc90b894261e8b088a82af210d7822d5b9092a1b5d169