AI Engineer, Intern

Postman•San Francisco, CA

2d•Onsite

About The Position

Postman is seeking an AI Engineer Intern for its Summer 2026 internship to work with the AI team on large-scale AI and agentic systems, from data pipelines through production deployment. The intern will own well-scoped pieces of real systems with mentorship from senior engineers. The role includes contributing to APIFlow-Bench, an open-source benchmark for API-development work, and helping develop an action-level AI safety benchmark in a simulated enterprise API environment. It also involves fine-tuning and evaluating open-weight models, running rigorous experiments, assessing low-bit quantized models, and helping build Postman’s in-product Agent Mode, primarily in TypeScript. The intern will document experiments, design decisions, and runbooks, and flag safety, fairness, or privacy concerns. The position is based in person at one of Postman’s listed office hubs. Postman states that roles based in the San Francisco Bay Area, Boston, Austin, New York City, Tokyo, and London work in office five days a week. Bangalore roles currently work in office three days a week and are expected to transition to five days by year-end.

Requirements

Currently pursuing a BS, MS, or PhD in Computer Science, Data Science, or a related quantitative field.
Hands-on experience training or evaluating machine-learning models; experience may come from course projects, research, hackathons, or a prior internship.
Solid Python fundamentals, including data structures, functions, and basic testing, and comfort writing and reviewing code outside notebooks.
Working knowledge of at least one deep-learning framework; PyTorch is preferred.
Clear written and verbal communication skills and a habit of documenting work.

Nice To Haves

Experience fine-tuning open-weight LLMs using SFT, LoRA, reinforcement learning, or distillation, with improvement measured on a benchmark.
Experience building LLM agents involving tool calling or multi-step loops, or building LLM evaluation harnesses or benchmarks and reporting results with statistical rigor.
A track record of shipping software end-to-end, including APIs and services, CLIs, Docker, CI/CD, or cloud; public GitHub code is a plus.
Interest or experience in AI safety and robustness, such as red-teaming, prompt injection, agent security, fairness, or interpretability.
Exposure to model-efficiency work such as quantization, low-bit inference, or serving optimization.
Evidence of rigor and initiative, such as publications, technical blog posts, ablation studies, or self-driven side projects with quantified results.
Fluency with AI coding tools such as Claude Code, Cursor, or Codex, while maintaining a deep understanding of the systems built.

Responsibilities

Contribute to APIFlow-Bench by designing and reviewing benchmark tasks and mock API environments, extending the Python evaluation harness and task-generation pipeline, and helping maintain the public multi-model leaderboard with statistical confidence intervals.
Help build an action-level AI safety benchmark, including scenario design, threat modeling for prompt injection, data exfiltration, and permission overreach, and auditable evaluation design.
Fine-tune open-weight models for tool-calling and agentic tasks using supervised fine-tuning, distillation, and reinforcement learning with PyTorch and open-source training tools on managed platforms and self-managed cloud GPUs.
Design and run rigorous experiments; evaluate training runs on benchmarks; support ablation studies and error analysis; track experiments; and report results, including cost.
Evaluate ultra-low-bit quantized models for on-device use by extending quantized-versus-full-precision benchmark comparisons and analyzing differences.
Help build Postman’s in-product AI agent, Agent Mode, using a minimal architecture with direct LLM API calls, tool loops, multi-step execution, and checkpointing, primarily in TypeScript.
Read open-source agent-harness source code and turn findings into design specifications and prototypes.
Document experiments, design decisions, and runbooks, and flag safety, fairness, or privacy concerns in model or agent behavior.