Fullstack Engineer (Data Platform)
Posted Updated
Summary
Infrastructure engineer at Medal.tv / General Intuition who owns the platform between billions of gameplay clips and in-house AI models: preprocessing pipelines, multi-region Kubernetes GPU clusters, storage/IO, and inference serving. Stack is Python, Go, Rust/C++, Terraform across multiple cloud providers.
- Billions of gameplay clips a year come in on one side. Large action models and world models train and serve on the other. Everything in between - the pipelines that turn raw footage into training data, the clusters that consume it, the storage and I/O that keeps them fed, the runtime that serves the results - is infrastructure, and it is what you own
- This is deliberately not a narrow role. We are not hiring a Kubernetes specialist, or a data engineer, or an inference person. We’re hiring someone who can move from cluster scheduling to disk throughput to a preprocessing pipeline to inference latency in the same week, and who becomes the technical reference other engineers bring their system designs to, across both GI and Medal
- You’ll work directly with the founding team, at a company small enough that the decisions are yours to make
- Our Stack:
- Kubernetes, multi-region. GPUs across cloud providers. Python and Go, with Rust and C++ where performance demands it. Terraform. In-house frontier models: action models, world models, video understanding
- You optimize inference in production - batching, quantization, KV cache, serving runtimes - and you own the latency and cost numbers rather than reporting them
- What We’re Looking For:
- You build the preprocessing pipelines that turn a very large corpus of raw gameplay video into training-ready data, at a throughput that keeps training from waiting on data
- We weigh two routes in the same. Either you spent years deep in infrastructure at a large tech company or a serious lab, then left to build your own thing as founder, co-founder, or founding engineer, and you’ve been at it for at least a year
- You own orchestration and GPU clusters - scheduling, utilization, capacity. Expensive hardware sitting idle is your problem, and so is a training run blocked behind the scheduler
- You treat disk and network I/O as a first-class constraint rather than an afterthought. At our data volumes it is frequently the bottleneck, and you know how to find out whether it is
- You can draw a circle around something substantial and say: this was mine. You decided how it was built, you chose the technologies, and you carried it to production. Not “contributed to a team that” - you made the calls, and you have several examples
- You are at ease across cloud providers and comfortable owning infrastructure as code, multi-region deployment, and the reliability of everything above
- Or you’ve spent five or six years going deep on hard infrastructure inside a big company or lab, you own a system people have heard of, and you’re ready for a place where you decide what gets built
- Either way, you’re still writing code today and you want to keep writing it
- We are not dogmatic about any of this. If you think we’ve made the wrong call somewhere, that’s a conversation we want to have in the interview
