Information Security undergraduate at Donghua University, Shanghai.
I work on local LLM systems, evaluation tooling, small software products, and games. I keep the engineering notes, test protocols, reproduction evidence, and release artifacts that mattered while building them.
A local-first, game-inspired career asset manager for scoring, ranking, and role-specific loadouts.
Built with TypeScript and React, with an interactive demo, automated tests, bilingual content, and a public release workflow.
An engineering case study built around running local AI on consumer laptop hardware.
Covers runtime selection, GGUF quantization, model routing, reasoning policy, multimodal setup, benchmarking, and the trade-offs that showed up in actual use.
A living benchmark for comparing local LLMs under frozen, reproducible reasoning-budget protocols.
The evaluation workflow includes fixed test sets, blind judging, score locking, and reproducible experiment records.
I also spend time reproducing local-inference failures, tracing them into upstream projects, and validating fixes.
MiniCPM-V
#1153 ·
validation
Reported a MiniCPM-V 4.7 GGUF/chat-template failure and verified the updated official artifact across 36 recorded runs with no recurrence.
Hermes Agent
#96360
Most of this work has been around GGUF inference, long prompts, MoE behavior, chat templates, Windows runtime edge cases, and reproducible validation.
UE5 games · visual novel · WeChat Mini Program · HarmonyOS