Projects
Why I did it: To address repetitive Q&A and overnight unattended periods in community groups, I was invited to deliver a scalable QQ bot and to systematize the dev–release–ops lifecycle.
Intended impact: Establish a plugin-based command framework, role/permission controls, and monitoring/alerting, making it easy to integrate LLM capabilities later.
Why I did it: Building on my cloud computing coursework—and because my blog had been set up and maintained by others with limited control over changes and uptime—I decided to build an end-to-end site from scratch to solidify fully managed cloud and DevOps skills.
Intended impact: Establish Infrastructure as Code (IaC), CI/CD, and least-privilege IAM templates.
Resume one‑liner: Built a cloud‑native site from scratch (S3/CloudFront + API Gateway + Lambda/FastAPI + RDS/DynamoDB); SLO ≥ [99.9%], P95 < [200 ms], rollback ≤ [10 minutes].
Suggested metrics: Availability, error rate, change failure rate, MTTR, LCP/TTFB, cost per request, automation coverage.
Why I did it: To provide my advisor’s project with reproducible experiments and a reliable data pipeline; I used a lightweight, interpretable SVM to validate signal effectiveness as a baseline for more complex models.
Intended impact: Prepare for quantitative roles while studying finance; deliver data cleaning/feature engineering/rolling backtesting and parameter‑grid scripts to ensure reproducibility.
Decentralized/Distributed Storage and Cloud Computing for AI Optimization (Overseas Collaboration)
Why I did it: My first overseas collaboration (University of California course practicum) to validate the feasibility of multi‑cloud fault tolerance and decentralized data processing, and to gain cross‑team collaboration experience and evidence.
Intended impact: Learn to balance consistency and cost; produce cross‑cloud data‑layout guidance and a disaster‑recovery runbook.
Why I did it: In competition settings, traditional D&A capability assessments were hard to compare and quantify; we needed actionable metrics and a scoring framework to support optimization decisions.
Intended impact: Build an Importance–Connectivity–Availability 3D model and recommendations; validate on [50] samples to improve interpretability and correlation.
Why I did it: Manual annotation of SEM images is slow and inconsistent; we needed an end‑to‑end segmentation pipeline to improve efficiency and objectivity.
Intended impact: Gain hands‑on familiarity with the AI development lifecycle and coding; deliver a cleaning–labeling–training–inference pipeline and evaluation benchmarks.