自研大模型应用分析师

Jul 15, 2026

Job Description

主要职责

1. 模型能力诊断与训练需求挖掘

  • 建立并持续维护模型能力评估矩阵(Capability Matrix),覆盖多轮对话核心维度:指令遵循、上下文一致性、Function Call 参数化与触发、知识检索与引用、语言切换、角色稳定性、安全合规等。
  • 基于线上 badcase、离线 benchmark、A/B 评测与业务反馈,开展系统性归因分析(Root Cause Analysis),区分模型能力缺陷、数据分布缺口、Prompt/工具链设计问题及场景定义偏差。
  • 将诊断结论沉淀为结构化训练需求(Training Demand Spec):明确问题定义、影响范围、严重等级、预期收益及优先级,纳入版本级需求 backlog。

2. 训练需求假设构建与实验验证

  • 对每一条潜在训练需求,构建可检验假设(Hypothesis):明确「若补充某类数据 / 能力,模型在何种场景、何种指标上应产生可观测改善」。
  • 独立设计并执行需求真实性验证实验,在投入大规模标注与训练资源前完成可行性闸门(Feasibility Gate),典型手段包括:
  • Prompt / 上下文消融实验:验证问题是否可通过工程侧调整缓解,而非必须训练解决;
  • Few-shot / In-context 探测:用小样本快速验证模型是否具备潜在能力、仅缺数据激活;
  • 对比评测(Side-by-side / A-B):量化当前版本与对照策略的差异,确认改进空间;
  • 反事实样本构造(Counterfactual Probing):检验模型在关键变量变化下的行为稳定性。
  • 输出标准化实验报告:含实验设计、对照组设置、变量控制、结论判定(Train / Prompt-fix / Reject)及后续行动建议,为算法与标注团队提供决策依据。
  • 对高优先级需求组织小规模 PoC 验证(Proof of Concept),跟踪实验结论与线上放量结果的一致性,持续校准需求判断准确率。

3. Feature 发现与数据需求规格化

  • 从业务场景、人机交互形态及竞品能力中,识别并定义自研模型智能交互新 Feature,输出 Feature Spec(能力边界、触发条件、交互范式、失败模式)。
  • 将通过实验验证的训练需求,翻译为可交付的数据需求文档(Data Requirement Document)
  • 参与 Roadmap 规划与版本对齐,确保 Feature 定义、benchmark 建设、训练集交付与模型发版节奏一致。

Requirements

任职要求

必备

  • 本科及以上学历,CET-6 及以上;语言学、心理学、人机交互、计算机、数据科学等相关专业或同等经验。
  • 2 年以上 NLP / 大模型 / 对话系统相关经验;有多轮对话、Function Call、RAG 或业务机器人落地经验者优先。
  • 具备假设驱动的问题分析能力:能从 badcase 中抽象可验证假设,设计对照实验并给出 Train / Prompt-fix / Reject 的明确结论。
  • 能独立撰写实验验证报告、训练需求规格书(Training Demand Spec)、数据需求文档(DRD),逻辑严谨、结论可复现。
  • 责任心强,抗压能力强,适应创业公司标准迭代快、跨团队沟通多的节奏。

About AI Rudder

AI Rudder is an AI product company transforming how business communicate to their customers.

We launched in 2019 with one of the world’s first enterprise-grade AI voice agents: built to hold real, intelligent conversations, not just play back scripts. Today, we serve over 500 enterprises across 20+ countries, processing 50 million calls every day and having delivered more than 8 billion calls to date. Our customers range from fast-growing startups to large enterprises.

When AI talks to people, it has to feel human. That’s what we build.

How We Work

AI-first: AI isn’t a tool we reach for. It’s the default we start from. Every team, every workflow: AI-first.

Global Collaboration: We operate globally and work as one team. Distance is a logistics problem, not a cultural one. We build the habits that make distributed teams actually work

Communication Principle: Simple, Trust, Transparent. We remove unnecessary complexity so people can focus on what matters most.

What We Offer

  • Learning & Development: Continuous learning through real projects, knowledge sharing session, mentorship, and cross-functional collaboration.
  • Performance Recognition: Regular performance reviews, feedback, and opportunities for career progression.
  • Global Exposure: Collaborate with international teams, global rotations, and work with global enterprise customers.
  • Meaningful Impact: Build AI solutions that create real value for businesses around the world.
  • Ownership & Autonomy: Freedom to take initiative and drive outcomes.
  • Team Engagement: Team Building, Local & Global Gatherings & Local Events Celebration

AI Rudder is an Equal Opportunity Employer. We celebrate diversity and are committed to creating an inclusive environment for all employees and applicants.

Apply for this role


Other News & Articles