Skip to content
View EMT5320's full-sized avatar

Block or report EMT5320

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
EMT5320/README.md

余凯琪，评测驱动的 LLM Agent 系统工程师

Email

生产经验

目前在某运营商安全研发团队做大模型算法与 AI 工程，主导和参与 15+ 个 AI 服务上线。SOC 多 Agent 研判系统扩展到 20+ 专业 Agent，代表性链路包含 240+ 次工具调用。系统规模上来后，上下文膨胀、工具预算和证据一致性成了最常处理的问题。我把这类问题拆成可运行的实验，确认边界后再把结果带回生产系统。

个人项目

项目 解决的问题 实现与结果
AlgoCoach-Flywheel 代码辅导的「教得好」能否像判题一样被执行验证？ 42 道 runner-ready 公开题 · 冻结 15 题 / 440 例评测基准 · provenance split 与泄漏审计 · 7B QLoRA SFT/DPO served eval 驱动 SFT 基线选择，并将增益与回归切片转入 regression mining
ContextGuard Agent Lab 任务、工具、预算固定时，四种 Agent 策略在哪一步分叉？ 16 scored cases × 4 strategies = 64 次受控运行 · 独立 grader · FastMCP 工具边界
Loomstead Agent 的每个动作能否回溯到动机、记忆与工具证据？ Godot + Python runtime · phase2.trace · counterfactual replay · audit packet
Tsukumo Agent 状态如何跨 runtime 交接，并保留来源、作用域与撤销语义？ Rust 5-crate workspace · receipt-first · 只读 re-entry · v0.1.0 · Linux/Windows CI

推理服务： LLM Inference Lab，面向 OpenAI-compatible endpoint 对比 QPS / TPS / TTFT 和延迟分位数，并区分历史压测导入与现场复跑。

技术栈

  • Agent 系统： 任务规划与 Tool Use / Function Calling、多 Agent 编排、LangChain / LangGraph、上下文工程、RAG / MCP、状态交接与 Trace / Replay
  • 评测与后训练： PyTorch / Hugging Face Transformers / PEFT、QLoRA SFT / DPO、benchmark、自动化 grader 与 LLM-as-a-Judge、训练 / 评测数据构建与治理、泄漏审计与人工校准
  • 推理与工程： Python、FastAPI / Pydantic、vLLM 多卡部署、OpenAI-compatible API、KV Cache / 长上下文与并发调优、Kafka、SSE / WebSocket

求职方向

LLM Agent 系统工程 / 大模型评测与后训练 / AI 应用工程 · 1908937833@qq.com

Pinned Loading

  1. algocoach-flywheel algocoach-flywheel Public

    Verifier-backed algorithm coaching system: practice runtime, provenance data flywheel, 7B SFT/DPO, and protocol-split evaluation.

    Python

  2. ContextGuard-Agent-Lab ContextGuard-Agent-Lab Public

    MCP-compatible agent strategy benchmark for tool traces, independent grading, and success-cost-context trade-offs.

    Python

  3. loomstead loomstead Public

    Agent Behavior Observatory: a Godot + Python multi-agent runtime with trace, counterfactual replay, eval, and audit artifacts.

    Python 1

  4. tsukumo tsukumo Public

    Receipt-first Rust runtime for portable agent state, trustworthy handoff, read-only re-entry, and an observable local TUI.

    Rust 1

  5. llm-inference-lab llm-inference-lab Public

    OpenAI-compatible endpoint benchmark for QPS, TPS, TTFT, P50/P95 latency, telemetry, and evidence-class leaderboards.

    Python