Staff ML Engineer - Efficient Production Inference

ATBF Labs; San Francisco (CA); Hybrid; USD 215,000 - 285,000; Full time

Benefits offered by this job

Equity

Health benefits

Job summary

ATBF Labs builds the inference engine for production AI. Every token a model serves in production runs through an inference stack emphasizing latency, throughput, cost, and quality. You will collaborate with the engine team to turn prototypes into production-grade systems, while benchmarks guide every decision.

The role emphasizes research with practical impact, including KV-cache compression, speculative decoding, and real-world deployment considerations.

Qualifications

• PhD in CS, ML, or related field or equivalent research experience.
• First-author publications at peer-reviewed venues.
• Hedge hands-on experience with LLM/VLM inference internals.
• Experience shipping research into production systems.

Responsibilities

• Conduct research on efficient inference: speculative and parallel decoding, quantization, sparsity, and KV-cache compression.
• Design, implement, and evaluate new methods against rigorous, reproducible benchmarks.
• Partner with the engine team to transition prototypes into production-grade systems.
• Analyze empirical results, find bottlenecks, and iterate quickly to improve model quality and speed.
• Track emerging work and bring high-impact ideas into our roadmap.

Skills

ML research

Empirical methods

Python

C++/CUDA

Experiment design

Communication with engineers/research

Fluency with ML frameworks

Education

PhD in CS/ML or equivalent

Tools

PyTorch

JAX

Job description

ATBF Labs builds the inference engine for production AI. Every token a model serves in production runs through an inference stack emphasizing latency, throughput, cost, and quality. You will collaborate with the engine team to turn prototypes into production-grade systems, while benchmarks guide every decision.

The role emphasizes research with practical impact, including KV-cache compression, speculative decoding, and real-world deployment considerations.