Recommendation
- 330,270 synthetic events across 15,000 users and 600 products
- Hybrid Recall@10: 0.2658
- Hybrid NDCG@10: 0.1374
Machine Learning Engineering
University of Washington · ECE + Data Science
Ziang(Andrew) Tan
Machine Learning · Multimodal Retrieval · Recommendation Systems · Computer Vision
Available June–September 2027
I build machine learning systems across recommendation, retrieval, computer vision, and deployment.
My recent work includes an e-commerce search and recommendation platform, tiny-object detection research, and real-time vision deployment on NVIDIA Jetson.
At a glance
Evidence over buzzwords.
Multimodal Retrieval
- Evaluated on 1,000 public Amazon Berkeley Objects products
- Image → product Recall@10: 0.8929
- Image → product MRR: 0.7811
Vision Research
- Failure-driven synthetic-to-real aircraft data selection.
- Recall@0.25 improved from 0.92% to 1.94% under a fixed label budget.
- High-resolution YOLO26n raised tiny-object recall by 10.19 percentage points.
About
ML rigor with hardware instincts.
I am pursuing dual bachelor's degrees in Electrical & Computer Engineering and Applied Mathematics: Data Science at the University of Washington, with expected graduation in December 2027 and a GPA of 3.80 / 4.00.
I like problems where modeling decisions, evaluation design, and systems engineering all matter. My embedded-systems background makes me especially attentive to latency, data provenance, failure modes, and the gap between an offline metric and a working deployment.
72Ksynthetic images in reproducible aircraft training workflows
330K+synthetic events in the recommendation benchmark
100+students supported as an Embedded Systems TA
Research & Experience
From controlled experiments to deployed systems.
Undergraduate Researcher — Tiny-Aircraft Detection
ARC Lab, University of Washington · Seattle, WA
- Designed failure-driven data selection for synthetic-to-real adaptation using fixed 300, 600, and 1,200-frame label budgets and controlled three-seed experiments tracked in Weights & Biases.
- Improved frozen-test Recall@0.25 on sub-15-pixel aircraft from 0.92% to 1.94% at the 1,200-frame budget while reducing false positives per negative frame by about 12% versus random sampling.
- Built a high-resolution YOLO26n tiny-object detector with P2 features, residual refinement, and channel-spatial recalibration and feature reinjection; improved sub-20-pixel recall@0.50 from 61.11% to 71.30%.
- Created reproducible training and evaluation workflows across 72K synthetic and 1,799 real images, with tiled inference and frozen-test safeguards.
Undergraduate Teaching Assistant
University of Washington · EE/CSE 474: Embedded Systems
- Support 100+ students in programming-intensive ESP32-S3 labs through code debugging, technical feedback, and weekly office hours covering FreeRTOS, synchronization, I2C, and hardware–firmware concepts.
Engineering Intern
Shenzhen Volcos Technology Co., Ltd. · Shenzhen, China
- Developed sensor-integrated embedded prototypes combining environmental sensing, motor control, ultrasonic sensing, and hardware/software debugging.
- Designed PCB layouts in Altium Designer and supported board bring-up, functional testing, and hardware validation.
Selected Projects
Recommendation, retrieval, vision, and deployment.
Search & Recommendation · May 2026–Present
E-commerce Multimodal Search & Recommendation Platform
Built an end-to-end e-commerce ML system combining personalized recommendation, multimodal product retrieval, and grounded product search.
Recommendation: Hybrid ranking reached 0.2658 Recall@10 and 0.1374 NDCG@10 on a frozen temporal test.
Multimodal retrieval: Image-to-product retrieval reached 0.8929 Recall@10 and 0.7811 MRR on 1,000 public Amazon Berkeley Objects products.
- Collaborative, content, and popularity candidate retrieval with hybrid ranking
- Pretrained SigLIP2 embeddings and FAISS exact search for text/image product retrieval, evaluated on 523 held-out image views
- GPT-4.1 mini tool-using search agent with BM25/vector routing, metadata filters, and grounded RAG; 36/36 expected tool routes on development queries, with 28 grounded responses, 6 abstentions, and 2 errors
- Docker deployment on GCP Cloud Run with GitHub Actions and OIDC; image size reduced from 4.62 GB to 1.80 GB
- Measured 708 ms client p95 and 258 ms server p95 over 200 sequential requests; recommendation evaluation uses a frozen 3,954-target temporal test and 20 automated tests
Edge ML & Robotics · Jun 2026–Present
Edge ML Robotic Fruit Detection & Autonomous Sorting
Built an edge perception and robotic sorting system on NVIDIA Jetson Orin Nano, sending vision-derived targets over TCP/IP to a Raspberry Pi-controlled mechArm with rescanning, placement verification, relocalization, and recovery.
Measured result: FP16 TensorRT improved learned-model throughput from 20.83 to 27.57 FPS; separate physical testing with the OpenCV/HSV backend achieved 18/18 grasps, 15/18 vision-verified placements, and 2/2 recoveries.
- 193 images and 346 audited bounding boxes with zero detected cross-split leakage
- PyTorch FP32 → ONNX → TensorRT FP16 on Jetson Orin Nano; separate OpenCV/HSV baseline: 52.39 FPS
- 10 formal physical scenarios with explicit frozen-test and deployment evidence
Robotics Modeling
Robotic Arm Kinematics & Trajectory Control
Implemented a Python workflow for 6-DOF manipulator modeling with Denavit–Hartenberg parameters, forward kinematics, numerical inverse kinematics, and pick-and-place trajectory generation.
View repository →
Embedded Systems
Autonomous Embedded Cat Care System
Built a real-time ESP32-S3 prototype with interrupt-driven ultrasonic sensing, hardware timers, buffered sensor logic, and state-machine motor control.
View repository →Technical Skills
A stack built for ML experimentation and deployment.
Programming
Python, SQL, C++, Java, JavaScript, C, Bash
ML / Data
PyTorch, Recommendation Systems, Information Retrieval, Multimodal Retrieval, SigLIP2, FAISS, BM25, RRF, Grounded RAG, LLM Tool Use, Weights & Biases, ONNX, Ranking, Model Evaluation, OpenCV, NumPy, Pandas, SciPy, model training, hard-example mining, domain adaptation, Ultralytics YOLO
Systems / Tools
Docker, GCP Cloud Run, GitHub Actions, OIDC, Linux, Git, Node.js, TensorRT, CUDA, TCP/IP, SSH, NVIDIA Jetson Orin Nano, Raspberry Pi, FreeRTOS
Methods
Temporal validation, leakage auditing, ablation design, ranking evaluation, reproducible pipelines, physical-system testing
2027 Internship Search
Seeking Summer 2027 MLE and SWE internships.
I am especially interested in recommendation and ranking systems, computer vision, edge ML, and roles where disciplined evaluation connects directly to product or physical-system behavior.
Recruiter Contact
Start a conversation.
Direct contact
If you are hiring for machine learning, recommendation systems, computer vision, or edge AI roles, send me a note.
- Available June–September 2027
- Based in Seattle, Washington
- Open to on-site, hybrid, and remote opportunities