高级AI工程师
Senior AI Engineer
关于该职位
我们正在构建能够从自然语言生成高质量运动的AI系统。该系统结合了前沿的语言模型、AI生成框架，以及专为结构化、可编辑动画设计的文本原生Motion DSL。
该职位涉及两个相互关联的领域：改进当前使用的生产生成系统，以及开发专门的模型，以更高质量、更低延迟和更好的成本效率直接生成Motion DSL。
你将处于LLM系统、训练后优化、代码生成、编译器、评估、数据工程和运动设计的交叉点。这是一个注重实践的工程角色，拥有端到端的所有权，并对产品有可衡量的影响。
主要职责
构建和改进生产生成系统
- 设计并推出在提示解析、模型协调、路由、检索、工具使用、结构化生成、验证、修复和视觉验证方面的改进。
- 诊断重复出现的故障模式，并将其转化为提示、数据、系统逻辑、约束或评估方面的持久改进。
- 构建基于编译器的反馈循环和确定性质量门禁，防止无效或低质量的输出到达用户。
- 开发实验和固定评估套件，以证明更改是否真正提升了输出质量。
训练专门的生成模型
- 为直接生成Motion DSL设计监督微调数据集、训练配方和训练后实验。
- 探索蒸馏、偏好优化、合成数据生成、强化学习方法和受限生成，当它们是合适的工具时。
- 使用稳健的评估选择检查点，涵盖正确性、视觉质量、可靠性、延迟和成本——而不仅仅是训练损失。
- 确定模型故障是通过数据、训练、推理、评估还是底层语言/运行时来解决最为合适。
构建数据和评估基础
- 使用过滤、来源追踪、版本控制、去重和污染控制，将生产生成转化为高质量的训练和评估数据集。
- 设计训练、验证和评估划分，以最小化泄露并保持有意义的泛化测试。
- 创建故障分类、硬负样本、回归套件和代表性提示套件。
- 将确定性检查、基于模型的评判者、渲染证据和人工审核结合起来，形成可靠的质量保障体系。
查看英文原文
About the role
We are building AI systems that generate production-quality motion from natural language. The system combines frontier language models, an AI generation harness, and a text-native Motion DSL designed for structured, editable animation.
This role spans two connected areas: improving the production generation system used today, and developing specialized models that can generate the Motion DSL directly with higher quality, lower latency, and better cost efficiency.
You will work at the intersection of LLM systems, post-training, code generation, compilers, evaluation, data engineering, and motion design. This is a hands-on engineering role with end-to-end ownership and measurable product impact.
Key Responsibilities
Build and improve production generative systems
- Design and ship improvements across prompt interpretation, model orchestration, routing, retrieval, tool use, structured generation, validation, repair, and visual verification.
- Diagnose recurring failure modes and turn them into durable improvements in prompts, data, system logic, constraints, or evaluation.
- Build compiler-backed feedback loops and deterministic quality gates that prevent invalid or low-quality outputs from reaching users.
- Develop experiments and fixed evaluation batteries that show whether a change genuinely improves output quality.
Train specialized generative models
- Design supervised fine-tuning datasets, training recipes, and post-training experiments for direct Motion DSL generation.
- Explore distillation, preference optimization, synthetic-data generation, reinforcement-learning approaches, and constrained generation where they are the right tools.
- Select checkpoints using robust evaluations across correctness, visual quality, reliability, latency, and cost - not training loss alone.
- Determine whether a model failure is best addressed through data, training, inference, evaluation, or the underlying language/runtime.
Build the data and evaluation foundation
- Turn production generations into high-quality training and evaluation datasets using filtering, provenance, versioning, deduplication, and contamination controls.
- Design train, validation, and evaluation splits that minimize leakage and preserve meaningful generalization tests.
- Create failure taxonomies, hard negatives, regression suites, and representative prompt batteries.
- Combine deterministic checks, model-based judges, render evidence, and human review into a reliable evaluation system.
What we're looking for
· Strong ML and software engineering
You have built and operated production AI or machine-learning systems, not only prototypes. You are comfortable moving across model behavior, data pipelines, APIs, infrastructure, evaluation, and product code.
· Hands-on LLM training experience
You have practical experience with supervised fine-tuning and modern post-training workflows. You understand how dataset construction affects model behavior and can explain how you prevent leakage, contamination, and misleading evaluation results.
· Strong evaluation instincts
You know that generative systems improve only when they can be measured. You can design experiments, regression suites, automated graders, and evaluation datasets that distinguish real gains from noise.
· Experience with structured or code generation
Experience with code-generation models, DSLs, grammars, parsers, compilers, structured outputs, constrained decoding, or program synthesis is especially relevant. The generated output is executable structured code, so syntactic and semantic correctness both matter.
· Production engineering judgment
You treat observability, reliability, latency, inference cost, caching, failure recovery, and maintainability as part of the ML system itself.
· Product and visual judgment
You can distinguish technically valid output from work that feels polished. Experience with animation, motion design, graphics, creative tooling, or multimodal systems is valuable, but not required.
Nice to have
- Experience fine-tuning or evaluating code-generation models.
- Experience with multimodal or vision-language models.
- Experience building model-based, human-in-the-loop, or rubric-driven evaluation systems.
- Experience with compilers, interpreters, language tooling, or program analysis.
- Experience with Rust, PyTorch, or distributed training infrastructure.
- Experience with preference optimization, reinforcement learning, or synthetic-data pipelines.
- Experience with animation, graphics, rendering, or creative software.
LottieFiles Perks
- Fully Remote Working Environment
- Flexible Work Hours
- A welcome gift and LottieFiles swag pack
- Bonus to set up your workstation at home
- Unlimited Leave Days
- Medical Insurance
- Generous learning budget
- Gym membership
- Co-working space membership
Please note: To proceed with your application, you must confirm your acknowledgment of this Privacy Policy by ticking the checkbox on the next page.
Read our Privacy Policy here:LottieFiles: Privacy Policy
Originally posted on Himalayas