> Markdown version of [/jobs/ext/3278712-founding-mid-training-rl-infrastructure-engineer](https://www.wearedevelopers.com/jobs/ext/3278712-founding-mid-training-rl-infrastructure-engineer). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# Founding Mid-Training/RL Infrastructure Engineer

- **Company:** PANO AI, INC.
- **Location:** Palo Alto, CA, United States
- **Experience:** Experienced
- **Contract:** Permanent contract
- **Skills:** Artificial Intelligence, Software Code Optimization, Profiling, Shard (Database Architecture), Software Debugging, Distributed Data Store, Distributed Systems, Language Modeling, Performance Tuning, Graphics Processing Unit (GPU), Large Language Models, Deep Learning, Data Pipelines
- **Published:** September 27, 2026
- **Apply:** https://www.indeed.com/viewjob?jk=010b257d955c327d

## About the Role

You will architect, scale, and tune the distributed and accelerator-aware training infrastructure for massive language models, including data pipelines, model sharding, checkpointing, rollout and reward-based RL, and overall system throughput. Experience with large-scale model pretraining and mid-training interventions is required.

Preferred experience with frameworks such as Megatron, Transformer-Engine, verl, slime, and related large-scale training and RL toolkits. Familiarity with training pipeline scaling, mixed-precision, memory optimization, and deep understanding of both supervised and RL-based training cycles is essential., * Significant experience with large-scale deep learning model training and distributed system design, including large GPU/TPU clusters (GB300/VR200, TPU v7x or similar).

 * Proven track record of pretraining and/or RL-based fine-tuning of large models (LLMs or comparable scale).
 * Deep familiarity and hands-on experience with frameworks such as Megatron, Transformer-Engine, verl, slime, or similar large-scale/foundation-model toolkits.
 * Experience with data pipeline design, model/data/optimizer sharding, checkpointing, rollout/reward design, and training operations at scale.
 * Strong accelerator (NVIDIA GB300/VR200 GPUs, TPU v7x) and memory optimization skills, including at cluster scale.
 * Excellent debugging, profiling, and performance-tuning abilities in distributed environments.
 * Comfort working in deeply technical, high-ownership, early-stage startup settings.

Cultural Fit

 * Hands-on technical excellence and strong engineering judgment
 * End-to-end ownership-from design to implementation to production
 * Bias for action: ship quickly, learn from failures, iterate
 * High intensity during critical milestones, focused on customer and product outcomes
 * Ability to work deeply and sustain high execution
 * Clear communicator, low ego, strong collaborator
 * Thrives in ambiguity, rapid change, and taking on multiple roles
 * Lifelong learner with a belief that capability compounds with time and effort

If you are excited to build the core training and RL infrastructure for Peano AI's vertically integrated foundation model-leveraging thousands of NVIDIA GB300/VR200 GPUs and TPU v7x accelerators, tackling the hardest scaling and optimization challenges, and helping push the boundaries of agentic AI-we'd love to talk.

## Description

* Build, optimize, and scale the distributed training infrastructure for Peano AI's foundation models, leveraging thousands of NVIDIA GB300/VR200 GPUs and TPU v7x accelerators for high throughput and efficiency.
 * Own and improve throughput, efficiency, scalability, cost, and reliability of end-to-end pretraining and RL pipelines.
 * Architect distributed data loading, sharding, checkpointing, batching, and accelerator utilization for multi-node LLM training at unmatched scale.
 * Integrate and optimize RL components: rollout, reward modeling, environment orchestration, and mid-training signal injections.
 * Work with and extend frameworks such as Megatron, Transformer-Engine, verl, slime, and other high-performance training libraries.
 * Tune memory usage, mixed-precision, and runtime performance for massive models and long-context workloads.
 * Debug and profile training performance bottlenecks-across model code, distributed compute, networking, and infrastructure.
 * Collaborate with application, infrastructure, and research teams to ensure our foundation models deliver on performance and functionality.
 * Translate research prototypes and experimental features into production-ready, scalable training systems that operate across state-of-the-art compute clusters.