AI Computing Research Intern
Naïve · US ·
- Work mode
- Remote
- Seniority
- Intern
- Employment
- Internship
Naïve · US ·
Research and ship systems optimizing AI agent inference for cost and speed using Python, PyTorch, and hardware acceleration techniques.
## About Naïve
Naïve is an autonomous company builder — we let businesses deploy AI employees to create/run entire companies. We're a $120M company backed by Y Combinator, Liquid2, DEEPCORE (Softbank), and more.
We're hiring interns who ship and hope to convert to full-timers.
## What You'll Do
- Research and ship the systems that make running thousands of AI agents dramatically cheaper, faster, and more reliable
- Optimize local / self-hosted model inference — quantization, batching, speculative decoding, KV-cache strategy, tensor & pipeline parallelism
- Build model routing that sends every request to the cheapest model that can actually do the job — frontier API when it matters, local when it doesn't
- Benchmark and deploy across hardware — GPUs, edge, on-prem, alternative accelerators — and turn the numbers into real deployment decisions
- Push on agent infrastructure: orchestration, caching, context management, and parallelization for fleets of concurrent agents
- Prototype recursive self-improvement loops — agents that improve their own tooling, prompts, and evals
- Own a research question end-to-end — frame it, run the experiments, ship the result into production
## The Role
You're not here to write papers nobody reads. You're here to find the cost/performance frontier and ship past it.
Every dollar and millisecond you save compounds across an entire fleet of AI employees. Our best interns take a benchmark on Monday and land a production cost win by Friday — and write the changelog entry themselves. This is research with a deploy button.
## Must-Haves
- High agency
- Strong systems + ML engineering — comfortable in Python and PyTorch, can profile, optimize, and ship without hand-holding
- Real understanding of how transformers actually run — attention, KV cache, memory bandwidth, throughput vs. latency tradeoffs
- Built and shipped something real — side project, OSS, hackathon win, research artifact, prior internship
- Comfortable with LLMs both as tooling and as objects of study — API calls, prompts, tool use, and what's happening under the hood
- Move fast, take feedback, push back when you're right
## Nice-to-Haves
- Hands-on with inference engines (vLLM, TensorRT-LLM, SGLang, llama.cpp)
- GPU kernel or low-level perf work (CUDA, Triton)
- Hardware benchmarking or deployment experience (cloud GPUs, on-prem, edge, alt accelerators)
- Built with agent frameworks
- Published, open-sourced, or blogged research/tooling
- Currently enrolled in a CS / EE / math program — or dropped out of one to build
*P.S. If you're serious about this role, send Sean a connection request with a note over LinkedIn.*