Member of Technical Staff, Post Training

Profound; New York (NY); On-site; USD 230,000 - 325,000; Full time

Job summary

Profound is seeking a Post-Training Engineer to transform proprietary answer-engine data into high‑quality training datasets and ML pipelines. You will own end-to-end work from dataset to reward design, training runs, and deployment, ensuring the model delivers accurate, production-grade results to Fortune 500 customers in a fast-paced startup environment.

You’ll work with Python, PyTorch, and RL methods, collaborating across teams to define evaluation metrics and ship improvements that matter

Qualifications

• Exceptional software engineering skills in Python and PyTorch (JAX a plus).
• Hands-on post-training experience (SFT, RLHF, DPO, GRPO, or similar RL methods).
• Experience with distributed training (e.g. FSDP, DeepSpeed, Ray; Kubernetes or Slurm).
• Strong instincts for evaluation, and the ability to notice when a metric is misleading.
• A track record of owning work end to end, from early experiment to shipped model.
• Strong opinions about how models behave, and you keep up with current research.
• You’d rather see your model in production than in a paper.
• Comfortable operating in a fast-paced startup environment with rapid iteration cycles.

Responsibilities

• Turn Profound’s answer-engine data into high‑quality training datasets.
• Build and iterate on SFT, preference optimization, and RL pipelines.
• Design reward signals that reflect what customers actually value.
• Train models for tool use and long‑horizon agentic workflows with the teams.
• Work with Evals to define what “better” means for each model.
• Prototype new methods on compute/data infrastructure and take to production.
• Ship what you train and monitor performance in production, feeding back into training cycles.

Skills

Python

PyTorch

JAX

SFT

RLHF

DPO

GRPO

RL methods

End-to-end ownership

Production deployment

Startup pace

Model evaluation

Tools

Kubernetes

DeepSpeed

Ray

SLURM

Job description

Profound is the fastest-growing AI marketing platform, used by 15% of the Fortune 500 including Comcast, Estée Lauder, and Walmart, and innovators like Ramp, MongoDB, and Figma. It is where marketing teams do the work: research, briefs, content, campaigns, reporting, all in one place, alongside agents that carry the load with them.

Built on 2B+ real user prompts and growing, Profound’s data shows you what people ask AI and what the models say back. Profound's AI Marketer joins that data with deep context about our customers to find the opportunities worth acting on, and then puts agents to work: automating reporting, drafting content, auditing pages for readability, running competitive research. It measures what works and gets better every day.

As a Post-Training Engineer, you will turn Profound’s proprietary answer-engine data into models that are measurably better at the work our customers rely on: analyzing AI visibility, attributing citations, classifying sentiment, verifying claims, using tools, and running long-horizon agentic workflows. Most post-training is measured against benchmarks. Ours is measured by whether a Fortune 500 marketing team gets the right answer. You’ll own the full loop, from dataset to reward design to training run to deployment, then watch how your model performs with real customers and feed that into the next cycle.

What you’ll do

• Turn Profound’s answer-engine data (prompts, responses, citations, query fanouts) into high‑quality training datasets

• Build and iterate on SFT, preference optimization, and RL pipelines

• Design reward signals that reflect what customers actually value: reward models, LLM judges, and verifiable rewards

• Train models for tool use and long‑horizon agentic workflows with the teams building AI Marketer and our analytics products

• Work with the Evals team to define what “better” means for each model, and hold every training run to it

• Prototype new methods on our compute and data infrastructure, and take the ones that work to production

• Ship what you train, own its performance in production, and turn regressions and customer behavior into the next training cycle

Who you are

• Exceptional software engineering skills in Python and PyTorch (JAX a plus)

• Hands‑on post‑training experience (SFT, RLHF, DPO, GRPO, or similar RL methods)

• Experience with distributed training (e.g. FSDP, DeepSpeed, Ray; Kubernetes or Slurm)

• Strong instincts for evaluation, and the ability to notice when a metric is misleading

• A track record of owning work end to end, from early experiment to shipped model

• Strong opinions about how models behave, and you keep up with current research

• You’d rather see your model in production than in a paper

• Comfortable operating in a fast‑paced startup environment with rapid iteration cycles

Nice to have (optional):

• Experience with agentic or tool‑use training, synthetic data, or reward model design

• Experience with retrieval or search‑grounded models

• Publications at venues like NeurIPS, ICML, ICLR, MLSys, ACL, or EMNLP

Location

This is an on‑site role based in our NYC or SF office, designed for builders who thrive on speed, iteration, and meaningful impact.

For this role, the expected base salary range is $230,000 to $325,000. Profound’s total compensation package is designed to be competitive and includes base salary, equity, and a full range of benefits and perks. Final compensation will depend on factors such as your skills, experience, qualifications, and location, and will be determined during the interview process. Our recruiting team will share more details about the full compensation package and benefits as you move through hiring.

#LI-PRO

Note: All official communication from Profound will come from a @tryprofound.com email address. If you're contacted by anyone using a different domain, please disregard and report it as spam.