AI Training and Prompt Engineering Specialist
约兰达是我们的人工智能保理。在 YO 平台的 V3上，约兰达就是接口。用户与她交谈而不是通过菜单打。她回答有关计划，数据，加装和 YOYO$余额的问题，并对用户账户采取行动。当她错误时，用户卡住了，产品失败了。她的知识是 Yo 生态系统，没有其他东西。当问题落后时，她把它转移到 YOnC,我们的计算合作社，将其导向专业模型。知道该线位于哪里，并保持它，是工作的核心。她遵循一个金色规则：在五个回复中，她把产品或服务放在用户面前。不是5个回应是有用的。五个回应到达某个地方真实。这个角色使她正确了。你写出指令来控制她的想法。你构建评估证明她工作。你写下训练数据来纠正她当她没有做的时候。你发送运行所有代码.YOlanda 在我们的自己的 LangChain 和 LangGraph 构成上运行。你将与拥有它的人两个工程师一起打开拉动请求对该图表。你不会管理任何人。你会每天都在模型里。谁这就是为了你故意破碎东西，然后解释他们为什么破解了。你可以读出一个错误的答案，然后用一句话说什么是错误的。你宁愿运行五十个测试，而不是赢得一个争论，即提示是更好的。你有意见关于语音和你可以捍卫它们。你写得足够好让人们注意到。你不。第一版本永远不会是你发送的。位置：远程，全球工作格式：远程；全职 WHAT YOU'LL DO · 拥有 Ylanda 的系统提示 人格，调度，安全规则，她拒绝了什么 以及每个用户行动背后的提示：计划改变，加上，平衡查询，报价，支持。· 拥有边界：· 建立一些拍摄库，管理文本，这样她永远不会在空中提示上打开或重复已经回答的问题。· 构建评估集：每个任务都得到了测试案例，通过条件和一个分数你可以跟踪发布的释放。如果得分下降，什么也没有运输。· 排名和批评输出 (RLHF),重新写弱答案到参考/细节调整数据，并检查计划价格，平衡和实况。
查看英文原文
ABOUT THE ROLE
House of YO is a connectivity and distribution business with compute layered on top. YOlanda is our AI concierge.
On V3 of the YO platform, YOlanda is the interface. Users talk to her instead of tapping through menus. She answers questions about plans, data, top-ups and YOYO$ balances, and she takes the action on the user’s account. When she is wrong, the user is stuck and the product has failed.
Her knowledge is the YO ecosystem and nothing else. When a question falls outside it she hands off to YOnC, our compute cooperative, which routes it to a specialist model. Knowing where that line sits, and holding it, is central to the job.
She works to one golden rule: within five responses she has put a product or service in front of the user. Not five responses to be helpful. Five to arrive somewhere real.
This role makes her right. You write the prompts that govern how she thinks. You build the evaluations that prove she works. You write the training data that fixes her when she does not. You ship the code that runs all of it.
YOlanda runs on our own LangChain and LangGraph build. You will open pull requests against that graph alongside the two engineers who own it.
You will not manage anyone. You will be in the model every day.
Who this is for
You break things on purpose and then explain exactly why they broke. You can read a bad answer and say in one sentence what is wrong with it. You would rather run fifty tests than win one argument about which prompt is better. You have opinions about tone and you can defend them. You write well enough that people notice.
You are relentless. The first version is never the one you ship.
Location: Remote, global
Work format: Remote, full time
WHAT YOU'LL DO
- Own YOlanda's system prompt — persona, tone, safety rules, what she refuses — and the prompts behind every user action: plan changes, top-ups, balance queries, offers, support.
- Own the boundary: decide what's inside YOlanda's frame of reference and what hands off to YOnC, and make the handoff seamless — same conversation, same voice.
- Build the few-shot library and manage context, so she never opens on a blank prompt or repeats a question she already answered.
- Build the evaluation set: every task gets test cases, a pass condition and a score you can track release over release. Nothing ships if the score went down.
- Rank and critique outputs (RLHF), rewrite weak answers into reference/fine-tuning data, and fact-check plan prices, balances and YOYO$ maths — the hallucination target on money is zero.
- Red team her: adversarial prompts, prompt injection, jailbreaks and abuse of the top-up and rewards flows.
- Write Python to run and score evaluations at volume, work in the API (streaming, function calling, structured output), and keep prompts versioned in git.
WHAT YOU NEED
- 2+ years building with large language models in production. Using a chat window every day is not this.
- Python you can ship: scripts, API clients, data handling, tests.
- LangChain and LangGraph — you can read a graph you didn't write and change it without breaking it.
- Retrieval in production: chunking, embeddings, reranking, grounding, and knowing when retrieval is the wrong tool.
- Exceptional written English, and bilingual Spanish and English to the same standard — YOlanda has one voice that has to survive the crossing.
- Method: change one variable, test it, record it, move on.
Nice to have: fine-tuning (SFT, DPO, LoRA); evaluation frameworks (promptfoo, Braintrust, LangSmith, DeepEval); mobile, telecoms or fintech; editorial or linguistics training.
WHAT WE OFFER
- Top-notch products disrupting the world of the traditional media-services industry
- A brilliant, highly collaborative team with a shared drive to achieve goals
- Flexible working hours
- Corporate equipment for work
- Competitive salary
- Real opportunity for personal and professional growth
Originally posted on Himalayas