Yuliang Peng
Transport Policy (2026) Paper Published Quant.ai HF Live Space E-Commerce Vision Engine HF Space

Hi, I'm Yuliang Peng.

Actively Seeking Software Engineer & Machine Learning Engineer Roles

Computer Science at University of Maryland. Passionate about building scalable infrastructure, distributed systems, and exploring the mathematics of AI.

02. Industrial History

Work Experience

Software Engineer Intern

ICBC USA · Financial AI & Operations Jun 2026 — Aug 2026 | Los Angeles, CA
  • Integrated an enterprise AI/LLM platform into banking workflows, developing Python/SQL adapters, prompts, and validation logic for document processing, information extraction, analytics, and financial applications.
  • Developed AI-assisted workflows for document understanding and financial analysis, transforming unstructured inputs into summaries, extracted fields, and decision-support outputs.
  • Designed SQL-backed case-state and audit models for 200+ financial exceptions, capturing lifecycle transitions, failure reasons, reviewer actions, and compliance outcomes for end-to-end traceability.
  • Developed reconciliation and monitoring workflows for deposit and card-payment failures, reducing repeated case lookup and follow-up effort by 20% while improving operational visibility and auditability.

Software Engineer (Full-time)

ModeSens · Data & Vision Engineering Apr 2025 — Sep 2025 | Remote, USA
  • Engineered components of a serverless, event-driven ingestion platform using AWS Lambda, SAM, SQS, Redis, and Asyncio/HTTPX, implementing asynchronous processing, retries, deduplication, and failure recovery.
  • Implemented modular JSON-LD / DOM / API parsers covering 360+ merchant schemas, normalizing heterogeneous product, pricing, inventory, sizing, brand, and image data into catalog records.
  • Implemented a coarse-to-fine vision-matching stack using perceptual hashing, SIFT/FLANN, CIELAB/HSV features, and SSIM, improving match accuracy by 25% and reducing manual duplicate review by 80%.
  • Implemented Redshift-to-S3 archival and ETL components using Python, SQL UNLOAD, partitioned storage, and Redshift Spectrum, validating row counts and timestamps before warehouse cleanup.

Software Engineer (Full-time)

Bank of China · Transaction Processing & Reconciliation Jul 2024 — Apr 2025 | Los Angeles, CA
  • Developed Java/Spring Boot transaction-processing and reconciliation services for financial workflows, supporting status, retries, failures, settlement states, and downstream reporting.
  • Implemented PostgreSQL-backed lifecycle controls including idempotency keys and automated reconciliation, identifying duplicates, invalid transitions, and missing settlements.
  • Investigated production failures through logs and root-cause debugging; enhanced Grafana monitoring for latency, retries, reconciliation gaps, and failures, reducing incident-triage time by 30%.

Software Engineer Intern

All People Jan 2024 — May 2024 | Remote, USA
  • Built Node.js/Express APIs and optimized SQL queries for a marketplace platform, improving transaction performance by 15%.
  • Built Python ETL pipelines to aggregate user behavior data and support analytics features for recommendation workflows.

Software Engineer Intern

Air China Jun 2023 — Aug 2023 | Chengdu, China
  • Built Spring Boot dashboards and Python data-sync scripts for flight operations, supporting crew scheduling and maintenance logs.
03. Engineering Systems

Featured Projects

E-Commercial Platform - Multi-Modal AI System Engineering
Hugging Face Space • Multi-Modal AI System

E-Commercial Platform

FastMCP · ReAct Agent Swarm · Multi-Modal RAG · C++17 SIMD · SIFT / FLANN · AWS Redshift · 362+ Parsers
πŸ€— Hugging Face Live Space GitHub Repo
< 1 ms C++ SIMD 95.2% Precision FastMCP / A2A Standard 362+ Merchant Parsers

Production-grade multi-modal AI system engineering engine for luxury e-commerce catalog deduplication and visual authentication across 362+ global merchant platforms (Farfetch, SSENSE, Gucci). Implements an Autonomous ReAct Agent Swarm orchestrating automated schema parsing, multi-national size normalization (US/EU/UK/IT/JP), visual similarity verification, and SQL data lake ingestion. Features Model Context Protocol (FastMCP) and decentralized Agent-to-Agent (A2A) tool communication, sub-15ms Multi-Modal Product RAG combining dense vector embeddings with SIFT keypoint clusters, an ultra-fast compiled C++17 SIMD core (< 1 ms), and a 5-layer pyramid matching array (pHash/dHash coarse filters, OpenCV SIFT + FLANN KD-Tree vector alignment, CIELAB non-linear color ΔE, and SSIM error heatmaps), backed by automated AWS Redshift Spectrum Parquet time-partitioned archiving.

πŸ€— HF Space FastMCP / A2A ReAct Agent Swarm Multi-Modal RAG C++17 SIMD (<1ms) OpenCV SIFT FLANN KD-Tree pHash / dHash SSIM Heatmap CIELAB Color ΔE AWS Redshift Spectrum 362+ Parsers
Quant.ai - Low-Latency Market Replay & AI Platform
Hugging Face Space & Quant AI Platform

Quant.ai

C++17 · Python · LightGBM · LambdaMART · Gaussian HMM · Frontier LLM Agents
πŸ€— Hugging Face Live Space GitHub Repo

ML-driven market simulation & decision platform. Built a signal pipeline over high-frequency market & limit-order-book data, engineering Order Flow Imbalance (OFI), microprice, depth imbalance, spread, and volatility features for short-horizon prediction. Trained LightGBM and LambdaMART models for return prediction and cross-sectional ranking, implemented a 3-state Gaussian HMM for market-regime detection, and built an agentic LLM layer to transform unstructured market events into quantitative trading hypotheses.

πŸ€— HF Space C++17 Python LightGBM LambdaMART Gaussian HMM LLM Agents OFI & Microprice Limit Order Book pybind11
InferRoute - LLM Inference Gateway
Google Hackathon Award

InferRoute

Python · FastAPI · Redis · PostgreSQL · Docker
GitHub Documentation Hub

A multi-provider LLM inference gateway integrating OpenAI, Gemini, vLLM, and Ollama with request validation, streaming responses, timeout enforcement, credential isolation, and containerized deployment. Implements TCP Vegas-inspired adaptive concurrency control and sliding-window circuit breakers; sustained 45.2 RPS at 100 concurrent clients with 99.4% success and 120 ms p95 gateway overhead under provider failures. Evaluated on 3.2M WildChat conversations, reducing estimated API cost by 54.2% while retaining 98.8% of baseline quality.

Python FastAPI Redis PostgreSQL Docker Adaptive Concurrency Circuit Breakers LLM Gateway
TabIt AI Interface
Featured Systems Project

TabIt - AI Tab Manager

React · TypeScript · Chrome Extension MV3
Store GitHub Documentation

An intelligent web browser workspace extension that implements local semantic grouping of open sessions. Uses custom client-side TF-IDF embeddings and the Leiden clustering algorithm running within multi-threaded Web Workers to resolve similarity mapping across 100+ active browser tabs inside sub-100ms. Runs entirely on-device to enforce user data privacy.

React TypeScript Web Workers TF-IDF Leiden Clustering Manifest V3
03b. More Systems & ML Work

Other Engineering Projects

Sanguosha (三国杀) Noname Web Game

Sanguosha (三国杀) Game

Interactive turn-based card game deployed as a web application on Hugging Face Spaces. Implements state machine engine, hero skill pipelines, and AI decision logic.

πŸ€— HF Space JavaScript Game AI State Machine
ApplyPilot AI

ApplyPilot AI

An advanced browser-based AI job application agent. Scrapes page DOM structures, maps fields via heuristics, and leverages Gemini to draft short answers. Persists profiles locally.

React 18 FastAPI Gemini AI Manifest V3
Face Verification MLOps

Face Verification MLOps

Embedding-based face verification evaluated on the LFW dataset. Implements similarity scoring, confusion matrices, and ROC-based threshold calibration. Fully Dockerized CLI verifier.

PyTorch Docker ROC Calibration MLOps
CV & Tactile Depth Prediction

CV & Tactile Depth

Proposed a mathematical framework for sub-micron depth prediction from tactile sensor (DIGIT) images. Reconstructs surface topology using GMM color segmentation and panorama stitching.

Python Computer Vision CNN GMMs
Semantic Product Matcher

Semantic Product Matcher

Architected a high-concurrency ingestion crawler pipeline covering 30+ retailers. Built a similarity engine combining pHash and SSIM to reduce manual validation effort by 80%.

Python Selenium pHash SSIM
Terpiez AR Mobile Game

Terpiez AR Mobile Game

Location-based AR game captured virtual creatures in real-time. Features GPS coordinate tracking, interactive dynamic maps, and player session states using Redis.

Flutter Dart Redis Google Maps
Aviation & Low-Altitude Tourism Intelligence

Aviation & Low-Altitude Intelligence

Generative ABSA pipeline over 27K+ reviews using Llama-3-8B (0.82 F1). Built coarse-to-fine multimodal validation with CLIP & Qwen2.5-VL (0.84 consistency F1) and LLM-as-a-Judge guardrails (vLLM/FastAPI) reducing structural hallucinations by 92%.

PyTorch Llama-3 Qwen2.5-VL CLIP vLLM FastAPI
Adversarial AI Game Agents

Adversarial AI Agents

Programmed spatial gameplay agents. Implemented pathfinding (A*, DFS), adversarial decisions (Minimax with Alpha-Beta Pruning), and particle filters for state estimation.

Python A* Search Minimax Particle Filter
Neural Network from Scratch

NN from Scratch

Designed and built a zero-dependency deep learning framework from mathematical foundations. Implemented forward/backpropagation, activation layers, and optimizers (SGD, Adam).

Python NumPy Linear Algebra Optimization
04. Scientific Inquiry

Research & Publications

Peer-Reviewed Journal

International Cargo Charter Operations and Preighter Performances in the Context of COVID-19: A Case of China

Chuntao Wu, Xiaohe He, Yuliang Peng
Transport Policy · Vol. 184, Article 104173 · 2026
DOI: 10.1016/j.tranpol.2026.104173
Article Read PDF

Abstract: This research presents the first comprehensive empirical study of international air cargo charter networks and passenger-to-cargo ("preighter") operations under external system shocks. Leveraging a proprietary dataset of CAAC flight permits, we constructed undirected networks at the city level, representing over 80,000 annual flights.

Network Science Methodologies: We formulated the charter logistics system as an undirected graph and computed topological indicators across time-varying snapshots:

  • Average Degree & Path Length: Quantified the densification of hub-and-spoke models, proving structural stability and small-world properties.
  • Node & Edge Betweenness Centrality: Isolated critical routing bridges and tracked dynamic adjustments under intense route volatility.
  • Operational Flexibility: Identified the pivotal structural role of freight forwarders (charterers) who integrated truck-to-air transit to maintain supply-chain robustness.
UMD Research Project

Research Assistant

University of Maryland, College Park Nov 2025 — Apr 2026

GPU-Accelerated Spatiotemporal Event Camera Embeddings

Designed, implemented, and benchmarked event-driven machine learning models. Built custom CUDA kernels to perform high-speed, parallelized feature mapping on sparse spatiotemporal data streams generated by event-based cameras. By merging spatial convolutions with time-decay indices into fused SIMD kernels, achieved microsecond-latency ML inference on continuous high-rate sensory streams.

CUDA C++ PyTorch SIMD Parallelization Event-based Vision
05. Academic Path

Education

Master of Science in Applied Machine Learning

University of Maryland, College Park

Sep 2025 — Expected May 2027 · GPA: 3.8 / 4.0

Relevant Coursework
Data Structures & Algorithms Operating Systems Systems for ML Cloud Computing

Bachelor of Science in Computer Science

University of Maryland, College Park

Sep 2020 — May 2024 · GPA: 3.6 / 4.0 · Dean's List

Relevant Coursework
Machine Learning Computer Vision Robotics Deep Learning Data Structures & Algorithms
06. Core Strengths

Technical Skills

Machine Learning & AI

  • PyTorch
  • TensorFlow
  • LightGBM
  • LambdaMART
  • Gaussian HMM
  • Llama-3 & Qwen2.5-VL
  • CLIP
  • vLLM
  • CUDA
  • Hugging Face
  • QLoRA / PEFT
  • NumPy & Pandas

Backend, Infrastructure & MLOps

  • Spring Boot
  • FastAPI
  • PostgreSQL
  • Redis
  • Docker
  • AWS
  • Grafana
  • Kubernetes
  • POSIX Sockets
  • gdb & Linux perf
  • ASan / UBSan
  • CI/CD & Git

Languages & Spoken

  • Python
  • C++
  • Java
  • JavaScript / TypeScript
  • SQL
  • C & Rust
  • English (Fluent)
  • Chinese (Native)
  • Japanese (Native, N1)