Senior LLM Training Engineer (Post-training)

Senior LLM Training Engineer (Post-training)

仕事概要

In short

You'll make Recho's AI models smarter for our use cases through post-training. You'll design the training methods, the reward design that "grades" the model, and the training data yourself, and you'll own the model's intelligence itself. Your job is to optimize general-purpose LLMs for our use cases, and down the line, we plan to offer that performance externally as an API business.

• Team language: English (Japanese not required)
• 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…

• An ML engineer or researcher who has worked deeply on LLM training, especially post-training
• Someone who has applied SFT, preference optimization or RLHF to a production model and shipped it
• Someone with a CS or math background (MS/PhD, or equivalent industry experience) and a track record of conference publications or open-source work

About the role

At Recho, we don't use general-purpose LLMs as they are. We optimize them for our use cases through post-training and build them deep into our product. Your core mission is to design, implement and run training methods such as SFT, preference optimization and reinforcement learning (RL), and to own training quality for post-training of a single model.

You'll work at the core of what makes the model smart. That includes reward design (reward models, RLVR, LLM-as-a-Judge), designing training data quality and mixture ratios, and improving the post-training algorithms themselves (loss functions, KL penalty, advantage estimation, training stability).

We also plan to offer the performance we gain here externally as an API and grow it into a second pillar of our business. The scope includes post-training for agentic use cases such as long CoT and reasoning, tool use and function calling.

What this role is not

• Not an on-call or SRE role. SLA management, incident recovery and infrastructure maintenance for online services are not the focus.
• GPU cluster operations and general SRE/DevOps are handled by a separate team.
• Building the infrastructure that runs training and inference (rollout systems, weight sync, GPU allocation) belongs to our RL Infrastructure Engineer role. You'll own the algorithms, rewards and data, and partner with that team.
• Pre-training from scratch is not the main job. The focus is post-training.

What you'll do

1. Design, implement and run post-training

• Design, implement and run SFT, preference optimization (DPO, KTO, ORPO, SimPO, etc.) and online RL (PPO, GRPO, RLOO, etc.)
• Post-training for long CoT and reasoning, tool use, function calling and agentic use cases
• Improve post-training algorithms: loss functions, KL penalty, advantage estimation, training stability and sample efficiency

2. Design rewards

• Build reward models, RLVR (verifiable rewards) and LLM-as-a-Judge pipelines
• Detect, analyze and address reward hacking and reward over-optimization

3. Manage training data quality

• Generate synthetic data and manage data quality: mixture optimization, difficulty control, deduplication and benchmark contamination checks
• Generate and filter preference and synthetic data

4. Build evaluation in-house

• Design new internal evaluation sets that correlate with product quality
• Design and build our evaluation platform in-house, and design human evaluation, A/B tests and annotation guidelines

We're happy to go into more detail during the intro call and interviews.

Why this role

• Own the core of what makes the model smart. You'll own the algorithms, rewards and data that optimize general-purpose LLMs for our use cases, and you'll be responsible for post-training quality. You won't just use a platform; you'll design the model's intelligence itself.
• Build a second business pillar from scratch. We plan to offer the performance gained through post-training as an API and grow it into a second pillar of the business. Your work spans product optimization all the way to a new business.
• Run cutting-edge methods in production. You'll apply top-lab methods, such as online RL like GRPO, long CoT and reasoning, and agentic post-training, directly to a product that major enterprises run in production.

Tech stack

• Languages & frameworks: Python, PyTorch
• Training methods: SFT, DPO, KTO, ORPO, SimPO, PPO, GRPO, RLOO
• Rewards & evaluation: Reward models, RLVR (verifiable rewards), LLM-as-a-Judge
• RL & distributed training: verl, OpenRLHF, TRL, FSDP, DeepSpeed, Megatron-LM
• Rollout inference: vLLM, SGLang
• Experiment tracking & compute: Weights & Biases, etc.; Slurm, Kubernetes

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!

必須スキル

Must-haves

Language: Professional-level technical communication in English
Overseas applicants welcome: If you relocate to Japan, we provide work visa sponsorship and relocation support.

Post-training design & implementation (SFT, preference optimization, RL)

• Experience applying SFT and preference optimization (DPO, etc.) to a production LLM and taking it all the way to release
• Solid understanding of RLHF/RLAIF in both theory and implementation. You can explain PPO and GRPO at the implementation level, and you've designed KL penalties and dealt with reward hacking in practice.
• Experience improving and optimizing post-training algorithms yourself (loss functions, KL penalty, advantage estimation, training stability, sample efficiency)

Reward design & evaluation

• Experience with reward design and building reward models (reward model training, RLVR, LLM-as-a-Judge), and with detecting, analyzing and fixing reward hacking and over-optimization
• Experience defining evaluation metrics that correlate with product quality, beyond benchmark scores

Training data & analysis

• Experience designing and improving data pipelines on the premise that data quality and mixture ratios drive model performance (preference/synthetic data generation, filtering, difficulty control, deduplication, benchmark contamination checks)
• Experience quantitatively analyzing training data and model outputs, pinpointing the root causes of failure patterns, and turning them into data and reward improvements (error analysis, ablation design)

Distributed training & engineering

• Hands-on experience with distributed training in PyTorch (FSDP, DeepSpeed or Megatron-LM), including designing and debugging multi-node training yourself
• Strong, high-quality implementation skills in Python and PyTorch

歓迎スキル

Nice-to-haves

Broader models & methods

• Post-training experience with multimodal models (image, audio)
• Experience with Japanese-specialized models, tokenizer design or continued pre-training
• Knowledge of inference optimization (quantization, speculative decoding, KV cache optimization)

Large-scale operations & track record

• Experience stabilizing training and handling failures on large clusters
• Publications at major conferences (NeurIPS, ICML, ICLR, ACL, etc.), or a notable track record in open source or technical blogging

Languages

• Japanese or Chinese. Our team works in English, but Japanese (the company's working language) or Chinese (spoken by many of our colleagues) makes cross-team collaboration even smoother.

求める人物像

What we look for

• Finds root causes with data: You go beyond benchmark numbers. You pinpoint the causes of failure patterns through data and experiments, and turn them into reward and data improvements.
• Improves algorithms yourself: You don't just use existing methods as they are. You dig into loss functions and reward design to make them better.
• Works backward from product quality: You define evaluations that correlate with product quality, and design training from there.
• Bridges research and production: You take cutting-edge methods all the way to a production model release.

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.

応募概要

給与

Salary

Compensation is based on your current compensation, skills and experience.

• Annual salary: ¥10M–¥20M
• Salary review: Once a year, based on your contribution and company performance
• Stock option plan

勤務地

WeWork Hanzomon (Hanzomon PREX North, 2-3-2 Kojimachi, Chiyoda-ku, Tokyo)
Hybrid: mostly in the office, with some remote work. For candidates based outside Japan, remote arrangements are open to discussion.

雇用形態

Employment type: Full-time, permanent. Starting with a paid contract engagement first is also an option.

勤務体系

Working hours

• Discretionary work system (professional type under Japanese labor law)
• Deemed working hours: 9 hours/day
• Start and end times are up to you. As a general guideline, the team's core working window is 10:00–19:00, adjusted as work requires.

Scope of duties and work location

This disclosure is required under Japanese labor law.

• Duties: Design and development of LLM post-training (training algorithms, rewards and data) at hire; subject to change as determined by the company
• Work location: WeWork Hanzomon at hire (some remote work possible); subject to change as determined by the company

Holidays & leave

Weekends, Japanese public holidays, year-end/New Year holidays (Dec 29–Jan 3), special leave, and paid annual leave (10 days in the first year)

試用期間

Probation period: 3 months, with no change in terms

福利厚生

Insurance

• Full Japanese social insurance (health, pension, employment)
• Commuting allowance (up to ¥25,000/month)

Office

• WeWork Hanzomon, 2 minutes from Hanzomon Station, with a dedicated floor, focus booths and meeting rooms
• Free drinks, including coffee, soft drinks and alcohol
• High-quality office chairs
• MacBook Air and external monitor provided

Productivity

We build AI for our customers, so we work AI-native ourselves.

• AI tools support
  • Engineering roles: Top-tier plans for Claude Code and Codex
  • Business & PM roles: Team plans that let you share context across your team
• Amazon Business account for work-related books and equipment

Team & communication

Beyond project meetings and day-to-day conversations on Slack, we regularly set aside time for the whole company to share progress, discuss it together and get to know each other.

• Monday Kickoff: Every Monday, we share everything from work updates to little things going on in our lives.
• Monthly All-Team Meeting: On the first Friday of each month, each team shares its results from the previous month. Afterward, we get together over food.
• All Hands: Twice a year, we share company direction and our mid- to long-term vision, bringing the whole organization together.

Smoking

• Indoor smoking prohibited; outdoor smoking area available