Remote Senior Applied ML Engineer - Real-Time LLM Routing

Kimchi; España; Remote; EUR 100,000 - 150,000; Full time

Benefits offered by this job

Equity options

Remote-friendly environment

Learning budget

Annual hackathon

Team-building budget

Equipment budget

Extra days off

Job summary

Kimchi is seeking an applied ML Platform Engineer to own the routing model end to end: features, objective, training, calibration, and release cadence. You’ll treat it as a product, not a research artifact, shipping improvements with measurable impact.

You’ll build production-parity evaluations, convert fleet data into training signals, and optimize the objective as task success × token cost × wall-clock speed under real concurrency. Remote-first & global team.

Qualifications

• 5+ years in applied ML or ML research with real-time decision models.
• Strong Python and SQL at scale for millions of request-level records.
• Know the difference between model improvement and easier eval.
• Have opinions on LLM benchmarks and agentic harnesses.
• Understand serving realities: TTFT, decode speed, batching, concurrency.
• Instrument before you optimize, with error bars and all.
• Wide mandate excites you.

Responsibilities

• Own the routing model end to end: features, objective, training, calibration, and release cadence.
• Build the evaluation that gates changes: production-parity benchmarks and holdouts by task family.
• Turn fleet data into training signals with proper labeling and failure categorization.
• Model the objective the user experiences: task success × token cost × wall-clock speed.
• Run the online loop: shadow, A/B, full rollout, drift monitoring, retrain.
• Push the frontier with calibrated decisions and abstention under uncertainty.
• Collaborate with platform engineers on instrumentation and data path support.

Skills

Applied ML

Python

SQL at scale

Model routing

Evaluation design

Instrumentation

Tools

PyTorch

ClickHouse

LLM APIs

GitLab CI

Job description

Kimchi is seeking an applied ML Platform Engineer to own the routing model end to end: features, objective, training, calibration, and release cadence. You’ll treat it as a product, not a research artifact, shipping improvements with measurable impact.

You’ll build production-parity evaluations, convert fleet data into training signals, and optimize the objective as task success × token cost × wall-clock speed under real concurrency. Remote-first & global team.