Member of Technical Staff, Post-Training & Applied Research

San Francisco Tensor Company•San Francisco, CA

1d•$275,000 - $315,000•Onsite

About The Position

At SF Tensor, we're building the future of high-performance compute by rethinking and rebuilding the AI stack from the hardware to the cloud. Our Kernel Optimizer finds the fastest possible form for code on any vendor and cluster topology, and our Model Foundry manages runs, simplifies research, and moves workloads across clouds and chips. We are backed by prominent investors and are looking for individuals who believe in the necessity of compute advancements for AI progress. This role focuses on the modeling side of our enterprise offering, which promises to turn a customer's dataset into a specialist model in days. We achieve this through techniques like SFT, RL, DPO, and distillation. The infrastructure is exceptionally strong, enabling rapid model training. Experiments are managed through Model Foundry, providing a versioned and reproducible environment for running experiments on optimal hardware. This powerful engine has been used for post-training large language models, robotics models, and pre-training AlphaFold v3.

Requirements

Shipped post-trained models into production and can talk honestly about the tradeoffs
Hands-on depth across SFT and RL (DPO, GRPO, PPO or similar)
Ability to judge evaluation honestly: what to measure, what a result means and when a number is lying to you
Comfortable owning data: curation, filtering, labeling workflows and synthetic generation
Proficient in PyTorch or JAX
Willing to sit with customers' domain experts to turn their intuition into a reward function

Nice To Haves

Worked on RL infrastructure at scale: rollout engines, distributed training or throughput debugging
Experience in distillation, quantization and speculative decoding
Post-trained for agents and tool use
Been forward-deployed or has customer-facing engineering experience

Responsibilities

Own the post-training pipeline end-to-end: data curation, SFT, preference optimization, RL, evals, distillation and finally deployment
Design reward functions and RL looks along customer domain experts, who know the task inside-out but not our training stack
Build an eval harness trustworthy enough to make a ship/no-ship call within a short window, especially where the target is subjective taste rather than a scored benchmark
Structure and generate datasets, including synthetic data pipelines, from whatever the customer actually has
Distill specialist models down into smaller models
Drive the time-to-model, which means finding what's actually on the critical path and removing it, run after run
Embed with customers as a forward-deployed researcher, then hand the pipeline over cleanly when their team is ready to take over