Projects
Systems, products & research.
Backend services, data pipelines, and full-stack products I have built at work and on my own, plus research on LLM agents at Northwestern.
Engineering
4 projectsByteDance
- AI Agent Intern on the Dreamina & CapCut strategy team.
- Built a stateful workflow runtime with WAL-backed checkpoints and idempotent writes; event recall rose from 32% to 95% with zero duplicates.
- Agent-written briefs went directly to Zhang Nan, former CEO of Douyin Group & ByteDance China, now head of CapCut.
CrowdListen
- #2 Software Engineer, one of two engineers on the product.
- Shipped a full-stack platform where 60–90 second analysis jobs run on Redis-backed queues and horizontally scalable workers across 11 data sources.
- Owned deploys and reliability: Docker on AWS ECS/ECR through GitHub Actions, with integration and load tests for the failures we actually hit.
ScenePrep
- Solo project: an authenticated REST API, a durable PostgreSQL job worker, an installable CLI, and local and remote MCP, all on one set of services.
- Every write takes an idempotency key, revision checks reject stale updates, and row-level security isolates tenants.
- Paid model calls reserve budget in a transaction before dispatch and settle afterward, with no hidden retries.
Building in progress
- Building an AI visual memory tool for content creators.
- Uses VLMs to help reuse visual style and content patterns from past work.
- MVP in progress; stay tuned.
Research
7 projectsExperienceBot
- Undergraduate Researcher at Northwestern's Delta Lab (DTR).
- Agentic search and recommendation: turns ambiguous goals into structured queries, retrieves evidence, and reranks results.
- On a 50-query offline benchmark, +24% facet coverage and +18% nDCG@10 over single-shot LLM and semantic-search baselines.
Agentic Sensory Data QC
- Agentic AI Researcher at Northwestern's REAL Lab.
- Building an agentic framework for sensory data quality control.
- Stay tuned.
GMPNN-CS++
The proposed GMPNN-CS++ model introduces a dual-contrasting sampling approach and self-attention mechanism for DDI prediction, achieving 97% overall accuracy.
Regeneron Science Talent Search Scholar '25
Dandelion PDE Model
Finalist Award (Top ~7% globally) at HiMCM 2023. Constructed a system of four coupled PDEs modeling dandelion population dynamics with Fisher logistic growth, advection-diffusion, and Brownian dispersal.
HiMCM Finalist, top ~7% globally
MCM 2023 - Optimization
Global Meritorious Winner (Top ~10% among 11,296 teams). The only high school team to achieve this on Problem B. Built optimization models using simulated annealing, logistic models, and Euler Method.
Global Meritorious Winner, top ~10% among 11,296 teams
IMMC 2023 International
Advanced to Final Defense (Top 26 among 900+ teams) at IMMC Greater China International Round. Applied Entropy Weight and K-Means clustering for optimized land development strategy.
Advanced to Final Defense, top 26 among 900+ teams
IMMC 2023 National
Advanced to International Round at IMMC Greater China National Round. Applied logistic regression, decision tree, and random forest to classify lizards into 26 species.
Advanced to International Round