Open Role
CarbonForge - Software Developer, AI Inference
THE COMPANY CarbonForge is building the inference optimization layer that makes large language models faster and cheaper to run — with joules per token as a first-class metric alongside latency and throughput. We sit inside vLLM and SGLang and reshape what the compiler can see. Our scientific co-founders are Pierre-Luc Bacon (CIFAR AI Chair, Université de Montréal) and Christophe Dubach (McGill). We are seed-funded and moving fast. No research role. No PhD required. A merged pull request closes the loop faster than a credential.
THE ROLE
We are hiring a Software Developer who owns the full build–test–clean–ship cycle on our inference stack. You will write production code, author the benchmark harnesses that validate it, and maintain the quality of the codebase with the same rigor a Red Hat or Cohere engineer brings to a critical open-source project. This is a Maintainer-class role. You are not here to write features and move on. You will come back, fix what broke, write the tests nobody else wanted to write, and refactor the module everyone was afraid to touch. You care about code health the way an infrastructure engineer cares about uptime.
WHY CARBONFORGE
You will work directly with a technically exceptional CTO and scientific co-founders who have built and published at the frontier of ML systems. The problem is hard, the codebase is young, and your fingerprints will be on the architecture from day one. Energy efficiency in AI inference is not a feature — it is the business. If that framing resonates, we want to hear from you.
WHAT YOU WILL BUILD
- Inference serving components inside vLLM, SGLang, or TGI — batching strategies, memory management, attention optimizations - Automated benchmark harnesses that track latency, throughput, and energy (joules/token) across every release - Regression test suites and CI quality gates that catch performance degradation before it reaches production - Agentic AI workflow infrastructure — orchestration layers, tool-use pipelines, multi-step agent reliability testing - GPU profiling workflows using Nsight and NCU to identify and fix bottlenecks at the hardware level
WHAT WE ARE LOOKING FOR
- 3–6 years of production software development experience — inference, systems, or platform engineering - Demonstrated ownership of a codebase over time: you wrote tests, reviewed PRs with conviction, and refactored what needed fixing - Benchmark harness authorship — you have written the framework that measured whether something was actually faster - vLLM, SGLang, or TGI familiarity — you have read the code, filed an issue, or submitted a PR - GPU profiling with Nsight or NCU — you have used a profiler to explain why something was slow and then fixed it - Python fluency; C++ or Rust at the systems boundary is a strong signal - Open source contributions — merged PRs in projects with real users carry more weight than a list of technologies
WHAT THIS IS NOT
- A research role — we do not need a PhD or a paper
- A management role — you will be writing code every day
- A notebook culture — production systems, clean tests, and shipped code are the outputs that matter
STRONG ADDITIONAL SIGNAL
- Previous role at Red Hat, Cohere, NVIDIA, Meta, AWS, DeepMind, or a company with equivalent engineering rigour
- Experience building or testing agentic AI workflows: tool-use pipelines, multi-agent orchestration, LangGraph or AutoGen
- QA automation background applied to ML systems — you have caught a model regression before it shipped Contributions to vLLM, Triton, Flash Attention, llama.cpp, or comparable inference/kernel projects
Apply
Send a short note explaining why this venture, and link your LinkedIn or résumé.