Machine Learning Engineer

Building AI systems that are measurable, grounded, and useful.

I’m Jiawei Hu, an ML engineer and graduate student at Johns Hopkins. I work across reliable LLM systems, computer vision, and the engineering needed to move models into real products.

Explore selected work View GitHub

01 / Selected work

Systems built to be tested, not just demonstrated.

Two projects spanning production-oriented RAG and applied urban computer vision.

How can I reduce CUDA memory usage?

Grounded answer

Use gradient accumulation and mixed precision...

[C1] Exact source citation verified

LLM systems · Independent project

ResolveAI

A production-oriented RAG system for grounded engineering support and incident intelligence.

Built an end-to-end retrieval and answer pipeline with BM25, dense embeddings, reciprocal-rank fusion, cross-encoder reranking, strict structured output, citation validation, safe refusal, and deterministic fallback. Evaluation is treated as a release gate, not an afterthought.

1.00Recall@5

0.833Fact recall

1.00Citation grounding

Live demo Source & evaluation

Computer vision · Urban AI research

Baltimore Building Condition AI

A map-based research prototype translating street-level imagery and model runs into parcel-level building-condition evidence.

Combined vision-language models, spatial joins, view-cone filtering, and auditable DuckDB exports to explore building condition at city scale. The fourth iteration surfaces multiple BCI model runs in an interactive Leaflet map.

3,588Scored records

10Model runs

510Parcels in default view

Explore the map View repository

Building condition

BCI · 72.4Common12 AI 2PL

LowerHigher

02 / Experience

Applied research with real-world constraints.

Mar 2026 — Present

Johns Hopkins University · 21st Century Cities Initiative

Research Assistant · Computer Vision & Urban AI

Developing a parcel-level Building Condition Index from street imagery and LiDAR, with zero-shot and weakly supervised vision pipelines for identifying visible signs of property distress.

May 2025 — Sep 2025

University of Washington

Research Assistant · Computer Vision

Built crosswalk-detection models for accessibility auditing across 120+ municipalities. Fine-tuning YOLOv8, Mask R-CNN, and CLIP improved mAP from 0.71 to 0.84 on 5,000+ annotated satellite images.

03 / About

I care about the space between a promising model and a trustworthy system.

My work combines model development with the less visible engineering that makes ML dependable: data pipelines, evaluation, failure modes, APIs, and deployment. I’m especially interested in teams applying AI to complex, high-impact problems where evidence and reliability matter.

ML systems

PyTorch, TensorFlow, Hugging Face, Embeddings, RAG, LoRA / PEFT

Engineering

Python, SQL, FastAPI, Flask, Docker, REST APIs, Evaluation

Applied ML

Computer Vision, LLMs, Speech Processing, Geospatial AI

Education

Expected Dec 2026

Johns Hopkins University

M.S. Computer Science & Security Informatics · GPA 3.9 / 4.0

June 2024

University of Washington

B.S. Informatics · GPA 3.8 / 4.0