Senior Backend / ML Ops Engineer

Drafted's Values

We're a small team working fully in-person in San Francisco. We value high-ownership builders who want to be a part of a talented, highly motivated team. We're guided by the following values:

• Own the mission. We take agency, act like owners, and see problems through to real outcomes.
• Build in the open. We value direct feedback, fast learning, and growth through honest collaboration.
• Move with care and speed. We iterate quickly while staying deeply respectful of our teammates.
• Seek the why. We challenge assumptions, think from first principles, and never stop asking questions.
• Design for everyone. We believe anyone should be able to design and build a home they love.
• Solve what matters. We embrace hard problems and create new paths forward when none exist.

The Role

Example Projects

• Building parallel generation pipelines where multiple workers race to fill output slots, with dynamic filtering based on post-processing results. Implementing claim coordination to prevent duplicate work, fallback logic to use best-available generations when hitting retry limits, and caching mechanisms to reuse generations across jobs (same user regenerating with the same prompt).
• Developing coordination mechanisms for capacity-constrained pipelines where maximum concurrency is fixed (reserved GPU instances, instance quotas, API rate limits) and peak demand exceeds available capacity -- implementing backpressure, admission control, and retry logic to prevent overwhelming downstream consumers.
• Implementing timeout and cleanup policies that account for high variance of computational complexity (p99 is 10x p50) and variable parallelism where completion time depends on concurrent worker count, which fluctuates dynamically based on queue dynamics and capacity constraints, without being overly conservative or prematurely terminating legitimately slow work.

Ideal Experience

• Building and scaling GPU-based inference services, optimizing for both low latency and high resource utilization.
• Job orchestration and load balancing with parallel generations, heterogeneous resource constraints (GPU, CPU, I/O), and multi-tiered queues.
• Implementing observability for latency attribution and failure diagnosis for multi-stage, asynchronous, and cross-platform pipelines.
• Designing fan-out architectures where upstream job completion triggers multiple independent downstream consumers that have mixed criticality, with some consumers blocking and others best-effort.
• Familiarity with modern cloud infrastructure: managed databases, job queues, edge compute/CDN, and PaaS deployment platforms.

Desired Skillset

• Knowledge of training infrastructure, especially distributed GPU training across multiple nodes.