Generally Speaking
I'm a builder who communicates. I write code for autonomous agents, RAG pipelines, and full-stack AI systems — but I also know that the best architecture diagram is worthless if nobody in the room understands it. I'm a high-signal communicator: brevity, precision, and the occasional well-timed joke to remind everyone we're building software, not performing surgery. I've bridged gaps between engineers, business stakeholders, clinicians, and regulators — and I genuinely enjoy the translation work most people find exhausting.
My foundation is deliberately multi-disciplinary. I started in physics, earned a BEng and BS in food science & engineering at CAU and Cornell, and spent three years in life sciences operations managing global partnerships and clinical research pipelines. Thermodynamics, fluid mechanics, process control, regulatory compliance, cross-functional coordination — none of it was "tech," all of it was training. Controlling volatile biological systems taught me to think in systems, and managing stakeholders across continents taught me to communicate with precision. Both skills transfer directly to building deterministic software and shipping AI products on non-deterministic infrastructure.
Currently at Penn Engineering, I'm pursuing a master's in computer science with coursework in machine learning and distributed systems, while building agentic workflows, MCP integrations, and production-grade LLM applications. I also call myself a Humor Retriever — loyal to the joke, always chasing the laugh, and deeply committed to the belief that if you can't explain a complex model with a smile, you don't understand it well enough. If you want to talk hybrid search, RAG optimization, system design, or why most meetings could have been a well-structured email — my inbox is open.
Education
University of Pennsylvania
Master of Applied Science in Computer Science
Aug 2025 - May 2027 (expected)
- GPA: 3.9
- Thriving through learning, building and cooking…
- Courses: Software Development, Computer Systems, Data Structure & Algorithm, Machine Learning, Distributed Systems
Cornell University
Bachelor of Science in Food Science
Aug 2018 - May 2022
"I would found an institution where any person can find instruction in any study."
- MAGNA CUM LAUDE
- Member of CALS Honor Society
- Major GPA: 3.9; Minor GPA: 4.0
- Courses: Product Development, Spreadsheet Modeling, Market Research
China Agricultural University
Bachelor of Engineering in Food Science & Engineering
Aug 2018 - Nov 2022
- Dual Degree Program
- Top 1% student
- Courses: Principles of Mechanics & Engineering, Food & Nutritional Sciences, all bio & Chem & Phys
Experience
An AI-powered ambient scribe platform for paramedics, generating real-time clinical documentation from field audio using speech-to-text and LLM-based report generation.
| Architected dual-path transcription pipeline combining real-time WebSocket streaming (for on-screen feedback) with asynchronous batch diarization (for final PCR generation), then optimized the backend from blocking synchronous calls to async submit-and-webhook flows to eliminate timeout failures on 45–120 minute field recordings |
| Designed intelligent long-transcript handling for Claude-based PCR generation with automatic chunking, sub-report synthesis, and fallback strategies to handle transcripts exceeding LLM context windows while maintaining NEMSIS v3.5 clinical compliance |
| Implemented end-to-end HIPAA-compliant infrastructure: AES-256-GCM encrypted at-rest storage in Supabase, JWT-secured auth with biometric re-authentication, encrypted on-device buffering for poor-connectivity scenarios, and durable PCR persistence before audio deletion |
| Built production-grade resilience patterns across all third-party AI service integrations (AssemblyAI, Anthropic Claude): exponential-backoff retries, circuit breakers, graceful degradation, and cost-aware fallback routing (VAD-based alternatives when API quota exhausted) |
| Validated architecture decisions through comprehensive stress-testing on real multi-speaker 45–120 minute field recordings, uncovering actual bottlenecks (transcript length, encryption performance at scale) and eliminating unnecessary complexity from the design |
A Techstars-backed, SOC 2 Type II-certified AI procurement platform that helps mid-market manufacturers source smarter, manage POs, and scale without scaling headcount — with an AI agent that delivers insights through email and Microsoft Teams.
- Owned full product lifecycle for mobile app and agent interaction experiences, shipping features from design to deployment
- Built an autonomous email agent (classify → reply → dispatch) with sensitivity detection using TypeScript, DeepSeek, and Nylas API, integrated with the platform’s AI procurement assistant
- Engineered cross-platform Gmail/Outlook threading with adaptive rate-limited daemons for production reliability
- Designed a persona-driven integration testing sandbox with template resolution, PDF generation, and SQLite crash recovery for regression testing and training data generation
Selected as Accreditation Operations Support for the FIFA World Cup 2026 in Philadelphia, applying engineering precision to manage high-volume credentialing with zero errors. Coordinating complex logistics for global stakeholders under high pressure, drawing on three years of PM experience across international business environments. Serving as a multilingual point of contact for international participants, ensuring seamless access control and world-class operations throughout the host city.
A worldwide e-commerce and distribution company in biotech that connects global suppliers and clients.
- Directed business development, sales operations, and logistics, driving growth initiatives and improving operational efficiency
- Negotiated partnership agreements with 100+ global companies, contributing to 85% of total sales revenue
- Implemented email marketing automation with JavaScript and analyzed customer data using Python to optimize sales strategies
- Represented the company at major U.S. pharmaceutical and biotech conferences; trained staff on market trends and adapted strategies to industry developments
We help characterize the nutritional, genetic and hormonal determinants of mineral metabolism on maternal and child health.
- Initiated role through onsite internship (Jun – Aug 2022) and was invited to return as a part-time remote coordinator to continue supporting clinical studies
- Recruited, screened, and retained human subjects, ensuring compliance with clinical research protocols and IRB requirements
- Coordinated participant scheduling and visits; assisted with IRB submissions, protocol development, and SOP documentation
- Performed statistical data analysis of dietary recall datasets using Python, R, and JMP to support research outcomes
- Oversaw biospecimen handling, storage, and quality control; managed lab operations and collaborated closely with PI and study team
A global independent market leader in genomics, clinical studies and BioPharma Contract Development and Manufacturing
- Trained in Laboratory Information Management Systems (LIMS), Electronic Laboratory Notebooks (ELN), and Standard Operating Procedures (SOPs)
- Performed client sample testing and assay development in a BSL-2 environment, adhering to cGMP/GLP and USP guidelines.
- Operated, calibrated, and maintained HPLC, CE, osmometer, turbidimeter, and UV-Vis spectrophotometer
- Presented assay development progress and sample analysis results to senior management, aligning tasks with departmental goals
Projects
A terminal coding agent with supervisor-subagent architecture, on-demand skills, and OTLP observability — from-scratch rewrite of Pi on Mastra.
- Supervisor + 3 specialized subagents (reviewer, explorer, fixer) with tool-level access control — read-only tools injected at construction so reviewer/explorer physically can’t write or execute. LLM-driven routing with parallel delegation via backgroundTasks
- On-demand skills system — 4 lazy-loaded SKILL.md files extracted from system prompts, discoverable and user-extensible. BM25 full-text search, 15 workspace tools across 5 modules, plus custom web_fetch tool via Mastra createTool()
- OTLP observability with full-content spans — custom exporter producing traceId/spanId/parentSpanId hierarchy, every model_turn carries complete prompt + response, every tool_call carries full args + result. Verified on Logfire; compatible with any OTLP backend
- Mastra runtime with LibSQL memory (working, observational, semantic recall), multi-provider config, tree-structured sessions, and Catppuccin TUI — 3-package TypeScript monorepo, 23 tests
An LLM-powered autonomous email agent with a built-in persona-driven integration testing sandbox.
- Dual-architecture — Autonomous Agent classifies incoming mail (ignore / auto-reply / manager draft) via a DeepSeek-powered classify → reply → dispatch pipeline; Integration Testing Sandbox generates realistic multi-persona email threads across Gmail and Outlook for regression testing and training data generation
- Event-driven daemon system — dual polling agents (buyer / supplier) with adaptive rate-limiting that automatically back off under API limits and recover; SQLite-persisted crash recovery with resume-from-breakpoint
- Cross-platform threading — engineered Gmail/Outlook message ID resolution via recipient inbox polling to produce proper In-Reply-To headers across Nylas grant boundaries
- Interface-driven LLM layer —
ClassifierandReplyGeneratorinterfaces swap between rule-based (zero-dependency) and LLM-powered (DeepSeek via Anthropic SDK) in one line; persona-specific prompts with temperature-tuned diversity for training data quality - TypeScript + Node.js + SQLite — 12 runtime dependencies; OAuth flow for mailbox connection with grant type labeling (buyer / supplier / manager); zero-dependency PDF generation
A production-grade omnichannel analytics platform — modeled after a real e-commerce business — with a survival analysis churn model, multi-source revenue reconciliation, and a DCF LTV simulator built for non-technical stakeholders.
- Omnichannel revenue reconciliation — unified contribution margin ledger across Shopify, Amazon FBA/FBM, and 3PL freight
- Subscriber retention intelligence — regularized CoxPH survival model scoring every member with a Churn Risk Index and subscription runway
- Growth ROI simulation — 8-parameter DCF LTV calculator with break-even sensitivity analysis and margin-accretion signal
- Domain-agnostic architecture — the full pipeline (ingestion → features → survival model → drift monitoring) transfers to any subscription or repeat-purchase business
Dingle is a personal hybrid search engine that indexes your own documents and lets you search and query them using natural language — powered by hybrid BM25 + vector search and a RAG pipeline backed by Claude.
- Built a C++ HTTP server from scratch — raw POSIX sockets, thread pool, BM25 inverted index, no third-party libraries
- Hybrid retrieval — fuses keyword and hybrid search via Reciprocal Rank Fusion for results that outperform either method alone
- Full RAG pipeline — top-k document retrieval passed to Claude Haiku with prompt caching for grounded, cited answers
- Personal knowledge base — index your own .txt documents and query them like a second brain
A Chrome extension that transforms your AI chat history into an interactive thinking tree
- Visualizing thought drift - Detects topic shifts and maps them as branching nodes in real time
- On-device AI intelligence - Runs semantic embeddings locally via WASM, no data leaves your browser
- Frictionless interaction - Inline “Branch Here” buttons appear directly on AI responses
- Surgical tree editing - Drag nodes with magnetic snap, reparent entire subtrees, undo/redo 20 steps
- Zero-conflict injection - Shadow DOM sidebar lives alongside ChatGPT and Gemini without CSS clashes
- Local-first privacy - IndexedDB storage with no cloud sync, no telemetry, no accounts
Ding's NeuroAtlas v1.0 presents a high-fidelity, interactive 3D visualization of human neuroanatomy, reconstructed from raw MRI data. The application allows users to explore the internal structures of the brain through a "Glass Brain" explosion interface, facilitating both educational and research-oriented anatomical studies.
Oystraz is a gamified health tracking and life simulation app where your real-world choices directly influence your virtual character's performance.
- Gamifying health - Your real-world choices power a virtual character
- Providing a safe stress outlet - Prank your virtual octopus boss instead of real confrontation
- Anti-hustle philosophy - Our AI companion Pearl actively discourages overwork
- Making tracking fun - Ocean theme, pixel art, and playful interactions
A professional web-based application for comprehensive genetic variant analysis
- Architected multi-API integration system for genetic variant annotation with clinical interpretation
- Designed gene symbol search feature and ACMG-compliant pathogenicity visualizations
- Directed API selection strategy and performance optimization (95% faster with caching)
- Established comprehensive testing framework (24 test cases) and Docker deployment pipeline
- Balanced technical trade-offs between feature scope, user experience, and development timeline
Bioavailability of Docosahexaenoic Acid From Biofortified Chicken
- Participated in the Empowering U.S. Broiler Production for Transformation and Sustainability project with a grant from USDA
- Analyzed and interpreted fatty acids content in human studies using JMP and Excel
- Completed an oral presentation and poster session titled Bioavailability of DHA From Biofortified Chicken
Even the gut microbiota of honeybees is powerful.
- Our results illustrate the prevalence of strain-level variations in the metabolic potentials and the distributions of antibiotic-resistant genes in Chinese bumblebee gut bacteria
- Performed honeybee dissections, sample collection, PCR sequencing, and meta-analysis of sequencing data; contributing 30h/week for two years
- Portions of my work were incorporated into an additional publications on PNAS, contributing to the study of gut microbiota
More About Me
I thrive on structure and discipline, yet I’m easily won over by the simple things in life: quality sleep, a great soccer match, and the occasional windfall of free food.
Long before I immersed myself in science, my world was defined by the 19x19 grid of a Go board. As a semi-professional player, I became one of the top-ranked female players in my province by the age of 10, a milestone built on years of strategic thinking and relentless practice. When AlphaGo emerged, it didn't discourage me; instead, it shifted my fascination from the game to the machine. Transitioning into tech felt less like changing paths and more like continuing the same high-stakes match, just against a different kind of opponent.
To keep my mental focus sharp, I maintain a strict physical routine. I’ve traded high-impact sports for high-consistency swimming to preserve my joints, 25km freestyle every month. For me, swimming is the ultimate "low-latency" activity for sustained cognitive performance. I draw constant motivation from athletes like Fan Zhendong, Ke Jie，Zhang Weili, and Kylian Mbappé. Their resilience reminds me that dedication is a universal language, whether you're on the pitch or at a terminal.
I’m a staunch advocate for the "early to bed, early to rise" lifestyle, having maintained such a biological clock that hasn’t wavered for almost two decades. During my focused hours, I usually have Khalil Fong’s mellow tracks on loop. I’ve always admired his calm professionalism and kindness, qualities I strive to bring into my own development work.
Beyond the screen, I’m a certified BLS provider. Earning that certification was a grounding reminder that impact exists outside of grades and deadlines. It’s about being prepared to contribute when every second actually counts.
Cooking is my final ritual for balance. Having spent years in Food Science, I treat my body like a high-stakes lab experiment: alcohol is strictly blacklisted, while sugar, dairy, and caffeine are on permanent probation. Crafting healthy, grounding meals is how I stay connected to my roots. However, all my dietary principles vanish the moment someone mentions "free food". I have a secret, undeniable weakness for Basque cheesecake and triple chocolate cookies.
As I said, Soccer is the latest spark in my professional journey. This May, I am joining the Accreditation Operations Support team for the FIFA World Cup 2026 in Philly, bringing my obsession with process optimization to the world’s biggest stage. Having managed complex biotech pipelines and medical protocols, I treat high-volume credentialing like a mission-critical deployment: security standards are the source code, and workflow efficiency is the non-negotiable KPI. It is the ultimate collision of my technical discipline and my lifelong love for the "beautiful game." However, my professional composure is highly vulnerable to a perfectly executed bicycle kick; I might be verifying your identity, but my heart is definitely on the pitch.
It’s worth mentioning that this summer I’ve been trying both a strict low-GI diet and moderate-intensity aerobic exercise to improve my physique and overall well-being.