“EVA”

Prompt/Template CI & Regression Harness Builder

Productionizes prompt changes: versioned prompts, offline eval gates, online shadows, and rollback — prompt-as-code.

Desert Chase / Dune-Scale Action

Large-scale action with vehicles and dust physics; evaluates motion blur, speed ramping, and wide action legibility.

RAG / LLM-Application Architect

Designs a production retrieval-augmented generation system: chunking, embedding, hybrid retrieval, reranking, evaluation, and guardrails. (LLM-native dev category.)

RBAC / ABAC Authorization Engine

Designs and codes a maintainable permission system: role hierarchy, resource-level checks, policy evaluation, and tests — preventing authorization sprawl.

AI Agent / Tool-Use System Builder

Designs reliable LLM agents: tool contracts, planning loops, guardrails, state machines, and evaluation of multi-step tasks.

量化投资

As an expert in quantitative investment, your task is to interpret the following strategies, evaluate their approach, and provide detailed training to new employees. The strategy content is provided below, and I…

LLM Fine-Tuning / Model Customization Planner

Decides between prompting, few-shot, RAG, and fine-tuning; if fine-tuning, designs dataset, method (LoRA), and eval.

Enterprise AI Supply-Chain Security Assessment and Governance Program

用于对企业全部AI资产（模型、数据集、适配器、MCP服务器、插件与智能体工具）开展授权的全链路供应链风险评估；适用于新AI系统准入评审…

AI-SBOM and ML Bill of Materials Generation with Provenance Attestation

为模型训练与部署流水线设计AI软件物料清单（AI-SBOM/ML-BOM）的生成、签名与验真方案；适用于需要满足模型溯源与合规审计要求…

Training and Fine-Tuning Dataset Poisoning Audit and Sanitization

审计训练与微调数据集的来源完整性、投毒与污染风险，并设计清洗与持续监控机制；适用于数据集准入与训练前安全评审。

Fine-Tuned Adapter and Merged-Artifact Security Scanning Before Deployment

为LoRA/适配器、合并权重与微调产物建立部署前安全扫描与准入门禁；适用于开放适配器仓库的企业集中管控。

MCP Server and Tool-Invocation Supply-Chain Security Audit

审计模型上下文协议（MCP）服务器、工具描述与凭据配置的供应链风险，落实最小权限与调用审计；适用于智能体平台上线前安全评审。

AI Agent Plugin Marketplace Vetting and Sandbox Isolation Architecture

建立智能体插件/GPTs类市场应用的准入审查、分级与沙箱运行隔离架构；适用于企业插件商店或开放智能体平台的安全治理。

Direct and Indirect Prompt Injection Defense Architecture and Detection Engineering

设计直接与间接提示注入（含工具/文档/RAG返回内容）的纵深防护架构与检测规则；适用于LLM应用与智能体的安全开发生命周期。

RAG Knowledge Base and Vector Database Poisoning Defense

防护RAG知识库、连接器与向量数据库中的投毒与间接注入，建立文档准入、完整性与检索侧监控；适用于企业知识问答与智能体检索系统。

Model Extraction and API Distillation Abuse Mitigation

防护商用/自研模型API遭高置信度查询蒸馏与克隆，设计速率管控、异常检测与水印追踪控制；适用于大模型推理服务运营方。

Multi-Turn Agentic Red Team Exercise Design with Rules of Engagement

设计针对自主智能体的多轮授权红队演练（权限滥用、工具链越权、目标偏离），含交战规则、止损与人因评估；适用于高自治AI系统的…

Current state assessment

Active testing

Risk assessment