Shuyu Li

Systems & networks · Distributed Systems · TypeScript · Microservices · CI/CD

A self-hosted personal cloud platform that runs my résumé API, blog comments, file storage, a mail inbox, page hosting, an LLM (Large Language Model) gateway and about twenty other services from one 2 GB Linux server behind api.lishuyu.app.

• Rewrote the platform in TypeScript (Node.js 24): 26 services as independently deployable plugins on a ~1,200-line deploy-and-rollback runtime built on the Cordis plugin framework.
• Zero-downtime hot swap: isolated boot → health check → atomic traffic hand-over → drain → 60-second probation with automatic rollback to the last-known-good version above a 20% failure rate.

Code github.com/StevenLi-phoenix/api

Applied AI & agents · Multi-Agent · Generative AI · Sandboxed Execution · RAG Memory

A personal AI agent bot built on the Anthropic Claude API tool-use loop, with sandboxed code execution, hybrid retrieval memory, scheduled 'dream cycle' memory consolidation, and multi-channel messaging.

Outcome a Generative AI / Multi-Agent personal assistant with persistent memory at half the consolidation cost.

• Multi-Agent personal Generative AI bot on the Anthropic Claude API tool-use loop with parallel tool dispatch and per-provider fallback (Anthropic primary, OpenAI fallback).
• Per-session OS-level sandbox: macOS sandbox-exec (macOS process sandbox) with custom SBPL (Sandbox Profile Language) profiles or Linux bubblewrap (Linux process sandbox using kernel namespaces) with namespace isolation.

Code github.com/StevenLi-phoenix/Memoo

Machine learning · Time-Series Forecasting · Neural Networks · Signal Processing · SPIE Paper

A peer-reviewed research project from high school exploring whether combining classical signal-processing with a Neural Network would beat either approach alone for short-term temperature forecasting.

Outcome the hybrid achieved better short-term prediction accuracy than either constituent method alone, and the work was published as a peer-reviewed paper (SPIE Proceedings Vol. 12285, paper 122851C, doi:10.1117/12.2637187, co-authored with M. Zhang and Y. Wang).

• Peer-reviewed publication: SPIE Proceedings Vol. 12285, paper 122851C, doi:10.1117/12.2637187 (co-authored with M. Zhang and Y. Wang).
• Hybrid Deep Learning + classical signal-processing pipeline: CEEMDAN (noise-decomposition technique that splits a noisy time-series into structured components) → per-component Neural Network → recombined forecast.

Paper doi.org/10.1117/12.2637187

Machine learning · LLM Fine-Tuning · Mixture-of-Experts · LoRA · Distributed Training

A custom Generative AI / Large Language Model (LLM) Fine-Tuned to write Chinese xianxia / cultivation / fantasy web-novels in genre-appropriate prose, scaled up from my v1 14B model to a 35-billion-parameter Mixture-of-Experts (MoE) model.

Outcome a 35B MoE LLM that produces longer, in-style chapter completions and specifically eliminates v1's premature-end-of-sequence bug.

• Generative AI / Large Language Model Fine-Tuning at the 35-billion-parameter scale on a Mixture-of-Experts (MoE) base model using LoRA (Low-Rank Adaptation).
• Distributed Deep Learning training on dual NVIDIA H200 GPUs (Vast.ai hourly) via Hugging Face Accelerate + FSDP (Fully Sharded Data Parallel).

Write-up blog.lishuyu.app/posts/qwen35-35b-lora-sft-novel · Model huggingface.co/Steven10429/qwen35-35b-novel-gguf · Model huggingface.co/Steven10429/qwen35-35b-novel-lora · Dataset huggingface.co/datasets/Steven10429/chinese-novel-sft

Machine learning · Knowledge Distillation · Transformer · ONNX · Synthetic Data

A reproducible knowledge-distillation pipeline that turns a one-sentence goal into a tiny task-specific Transformer.

Outcome on support-ticket triage the student is 211× smaller (4.0 MB) and about 230× faster (0.51 ms on CPU) than its teacher.

• Goal-to-model pipeline: spec → synthetic data → teacher labels → tokenizer → train → calibrate → int8 ONNX export → eval → package, driven by one command.
• Support-ticket triage: 4.0 MB and 0.51 ms on CPU — 211× smaller and ~230× faster than the 850 MB ModernBERT teacher.

Code github.com/StevenLi-phoenix/laya-tiny

Machine learning · Medical Imaging · CNN · Kaggle · Deep Learning

A Computer Vision / Deep Learning project for SIIM-FISABIO-RSNA, a public Kaggle competition with thousands of teams worldwide: identify and localize COVID-19 indicators on chest X-ray radiographs.

Outcome top-tier final ranking — Silver Medal — out of thousands of competing teams, for combined accuracy across the classification and localization sub-tasks.

• Kaggle Silver Medal — top tier worldwide among thousands of teams in the SIIM-FISABIO-RSNA Computer Vision competition.
• Trained Deep Learning Convolutional Neural Networks (CNN) for both image-level COVID classification and lesion-level localization on chest X-rays.

Machine learning · Medical Imaging · Computer Vision · Deep Learning · Segmentation

An automated medical-imaging pipeline that outlines brain tumors directly on MRI scans, replacing a manual step that previously took radiologists several minutes per scan and produced slightly different boundaries each time.

Outcome tumor outlines now generated automatically in seconds per scan, with the same answer every run — radiologists move from manual tracing to lighter QC review.

• Fine-Tuned nnU-Net (a published medical-imaging Deep Learning CNN) on the lab's specific MRI sequences and tumor classes.
• Computer Vision pipeline: minutes of manual radiologist tracing per scan replaced by seconds of automated, repeatable Neural Network inference.

Systems & networks · Operating Systems · RISC-V · Networking · C

A from-scratch operating system for 64-bit RISC-V that boots in QEMU and performs a certificate-verified HTTPS GET to www.google.com, then grew into a small preemptive OS.

• Hand-written network stack from Ethernet frames up through TCP (handshake, retransmit, flow control, teardown), verified in Wireshark.
• TLS 1.3 HTTPS GET to www.google.com with full X.509 certificate-chain verification — mbedTLS wired to my own TCP, virtio-rng entropy and RTC clock.

Write-up blog.lishuyu.app/posts/手搓riscv-os-给google发一次验证过的https