Falcon Compute Collective is hiring a Staff Contract PyTorch Performance Specialist on the Performance team (Remote · United States). This is a hands-on role in applied machine learning—not a generic “AI enthusiast” post.
You will own modeling, evaluation, and production ML delivery end to end: problem framing, offline evaluation, safe rollout, and iteration from production signals. Success is measured with clear metrics (quality, latency, cost, or business KPIs)—not slide decks alone.
Falcon Compute — 10-week contract: graph breaks, compile modes, and CUDA kernels for recommendation towers.
Remote global (overlap EU or US).
Falcon Compute Collective needs someone who can move modeling, evaluation, and production ML delivery from experiment to production—with tests, monitoring, and stakeholder-ready communication.
The Performance group partners closely with product and domain experts; you will translate ambiguous requirements into measurable ML or LLM outcomes.
What you will do
- Design and ship modeling, evaluation, and production ML delivery with offline evals, dashboards, and rollback plans.
- Improve datasets, labeling, and metrics that correlate with user or business outcomes.
- Review model changes for safety, bias, latency, and infrastructure cost before wide release.
- Document decisions so on-call, support, and compliance teams can act without guesswork.
- Collaborate async with written specs, demos, and recorded design reviews.
What we look for
- Demonstrated production use of core skills such as Python, PyTorch, Docker, CUDA—with repos, writeups, or metrics we can verify.
- Comfort operating the listed toolchain (AWS) in real environments—not tutorials only.
- Ownership of at least one feature or model from design through post-launch review—including failure modes you handled.
- Clear communication on trade-offs (quality, cost, latency, safety) with product and non-technical stakeholders.
- Eligible to work in United States (or remote policy as stated in the listing).
Benefits & ways of working
- Paid weekly
- optional extension.