Senior LLM Data Engineer (Training Data & Evaluation)

■仕事概要
In short
You'll design and build the "curriculum" that makes Recho's AI smarter: the training and evaluation data for our LLMs, built end to end from real-world business data. Starting from call logs and our clients' business systems, you'll decide what the model should learn, build the data, and put systems in place to guarantee its quality. Today, data building is ad hoc and varies from project to project. Your job is to turn it into a repeatable, standardized workflow.
Team language:
English. Japanese at around JLPT N2 level is also required, since you'll work directly with Japanese call logs and business data.
Salary:
¥10M–¥20M per year, plus stock options
Location:
Tokyo (Hanzomon), hybrid. Candidates based outside Japan are welcome. We offer visa sponsorship and relocation support, and we're open to discussing fully remote arrangements.
You might be a fit if you are…
Someone who has built post-training data (SFT, preference, synthetic data) at a foundation model company
Someone at a customer service or conversational AI company who has designed training data from real business logs and systems
A data engineer or ML engineer who has designed and implemented data integrations with business systems yourself
About the role
At Recho, we believe the ceiling on model performance is set by the quality of rewards and data. Your core mission is to design and build LLM training and evaluation data end to end, starting from real-world data such as call logs, business systems, FAQs and manuals.
Right now, building data for each project depends on individual effort and isn't reproducible. You'll turn the whole process into a standardized system: cleansing, anonymizing and structuring raw data; building SFT and RL data (synthetic data generation, difficulty control, deduplication, mixture tuning); and maintaining evaluation sets and benchmarks.
This is the foundation for scaling to the next 20–30 client projects. You'll also design and run the "data flywheel" that turns error analysis of training results into better data. Down the line, this role could grow into leading our data and evaluation team.
What this role is
not
Not an on-call or SRE role. SLA management and infrastructure maintenance for online services are not the focus.
GPU cluster operations and general SRE/DevOps are handled by a separate team.
The final design of training algorithms and rewards (SFT/RL methods, loss functions) belongs to our LLM Training Engineer role. You'll partner with them on data and evaluation sets.
Defining client requirements and what a "correct response" looks like belongs to our Data PM role. You'll turn those definitions into real data.
What you'll do
1. Investigate, design and prepare real-world data
Investigate each project's business systems and data, and design training and evaluation data end to end
Cleanse, anonymize and structure real-world data such as call logs
2. Build training and evaluation data
Build SFT and RL data: synthetic data generation, difficulty control, deduplication and mixture tuning
Implement reward functions and evaluation sets, and maintain benchmarks for each project
3. Run the data flywheel
Design and run the loop that turns error analysis of training results into data improvements
Build data pipelines with reproducibility and versioning built in
We're happy to go into more detail during the intro call and interviews.
Why this role
Set the ceiling on model performance.
That ceiling is set by the quality of rewards and data. You'll decide what the model should learn and what belongs in the prompt, and build the foundation of its intelligence yourself.
Turn ad hoc work into a system, from zero.
You'll rebuild fragmented, project-by-project data work into a repeatable data flywheel, and design the foundation for scaling to 20–30 projects.
Work at the front line of real data and real products.
You'll work with actual call logs and business systems and turn real customer-service problems into data. This isn't clean research data; it's the data that runs production.
Tech stack
Languages:
Python (pandas, Polars, Spark, etc.)
Data infrastructure:
ETL, versioning, reproducibility, SQL, API and database integrations
Synthetic data & evaluation:
Synthetic data generation via LLM APIs, LLM-as-a-Judge
Training framework inputs:
TRL, verl (formatting SFT and RL data)
Our product
We build the Recho AI Voice Platform, which delivers voice conversations natural enough to be hard to tell apart from a human. We also deploy and operate Voice AI Agents for large enterprises.
Our edge is a platform for building and improving high-quality agents quickly. It captures industry know-how and turns user feedback into a fast, continuous improvement loop. Our agents now handle complex, large-scale contact center operations in finance, manufacturing and the public sector. In some deployments, they've resolved every call without handing off to a human agent.
Hiring process
Intro call → Case assignment & first interview → Paid work trial* → Final interview → Offer meeting

• The work trial is 1–2 consecutive days at our Tokyo office, working on real tasks alongside the team, and it's paid. It's a chance for both sides to check fit and get a feel for the work.

If you're based outside Japan,
we'll agree on the format together. We won't ask you to travel to Japan just for this step.
If your schedule or your current employer's side-work policy makes the trial difficult, just let us know.
We may also ask for references.
About Recho
Recho develops foundation models for customer-support AI agents and also deploys them for clients, all within one organization. Our core strength is voice AI technology we built entirely in-house, which resolves the industry's long-standing trade-off between quality and cost.
We've closed a Series A of approximately ¥700M and are rapidly rolling out to demanding production environments at major companies such as SBI Shinsei Bank and KDDI.
Learn more
Corporate website
Team interviews (note, in Japanese)
Company deck
Case study: Partnership with KDDI and Altius Link (Japanese)
Case study: SBI Shinsei Bank (Japanese)
2-minute voice demo
— give it a listen!

Must-haves
Language:
Professional-level technical communication in English, plus Japanese at around JLPT N2 level
Overseas applicants welcome:
If you relocate to Japan, we provide work visa sponsorship and relocation support.
Experience designing and building LLM training and evaluation data yourself, starting from business systems and real-world data
Large-scale data processing in Python (pandas, Polars, Spark, etc.), and designing and implementing data pipelines (ETL, versioning, reproducibility)
Experience quantitatively analyzing real-world data quality and turning noise, bias and missing data into improvements
Ability to design data by separating what the model should learn from what should be passed in the prompt
SQL, and integrating with APIs and databases
Experience generating synthetic data with LLM APIs and filtering it with LLM-as-a-Judge
Understanding of the input formats of training frameworks (TRL, verl, etc.), and ability to format SFT and RL data for them

Nice-to-haves
Data & methods
Building SFT, preference and RL data
Designing reward functions and evaluation sets
Anonymizing and managing data that contains personal information
Domain
Domain knowledge in customer service, contact centers or voice conversation

What we look for
Starts from "what should the model learn?":
You don't just collect and clean data. You think first about what the model needs to learn.
Finds the signal through careful analysis:
You don't mind the slow work of digging through real data, and you find what really matters in it.
Turns vague requirements into concrete data:
You talk with business-side stakeholders and translate fuzzy requirements into specific data.
Builds systems hands-on:
You build repeatable systems while getting your own hands dirty.
Our values
These are the values our team relies on to deliver Recho's value as fast as possible.
Exceed Expectations:
We keep exceeding the expectations of customers and teammates by delivering both speed and quality.
Team-First:
Each of us takes professional ownership, and we maximize the team's results, not individual ones.
Continuous Feedback:
We engage seriously and speak candidly for the sake of results. The harder something is to say, the more we say it with honesty, respect and care.
Earn Trust:
Trust is the foundation of every collaboration and every challenge. We act ethically and keep personal and company matters separate. Trust takes time to build and a moment to lose.