What are the responsibilities and job description for the Software Engineer - Machine Learning III position position at DGN Technologies?
Machine Learning Engineer, Prompt Safety & Agent Security
Lab Summary
The Developer Quality Innovation Lab at Samsung Research America builds the automation and tooling that powers data acquisition, safety, and evaluation for Samsung's mobile platform products. Our systems collect, curate, augment data and develop intelligent solution to protect models that fuels the foundation models and AI features shipping across Galaxy devices — and operate the evaluation pipelines that gate their quality before and after launch. We work closely with modeling, device, and product teams to close the loop from on-device signals and user feedback back into training data, faster and at higher quality.
Position Summary
We are looking for an experienced Machine Learning Engineer to lead the development of prompt injection and prompt safety models that protect Samsung's downstream agentic AI systems across phone, cloud, and XR/AR. You will design, train, and deploy classifier and guardrail models (both cloud-based and hybrid on-device) that screen agent inputs and outputs for injection attacks, unsafe content, and policy violations. A core part of the role is post-training these models with RLHF, DPO, and related optimization techniques to push detection accuracy and false-positive rates beyond what off-the-shelf solutions provide.
Role and Responsibilities
- Design and train prompt injection detection models and prompt safety classifiers that operate on both inputs to and outputs from Samsung's agentic AI systems.
- Build hybrid deployment pipelines that split safety inference between on-device (phone, XR/AR) and cloud, optimizing for latency, privacy, and detection coverage.
- Apply post-training techniques (e.g. RLHF, reward modeling, policy optimization) to optimize guardrail model performance, calibration, and robustness against adaptive adversaries.
- Curate and generate adversarial training data: direct and indirect prompt injections, jailbreaks, tool-use exploits, and unsafe-output cases drawn from red-teaming and production signals.
- Build evaluation harnesses that measure attack success rate, false-positive rate, latency, and on-device footprint across model iterations and threat categories.
- Partner with agent, device, and platform teams to integrate safety models into mobile-use agents, XR/AR assistants, and cloud agentic workflows, and to close the loop from production incidents back into training data.
- Work cross-functionally with security researchers, modeling teams, and product engineers; document methods and, where appropriate, contribute to patents and publications.
|
Required Qualifications
Preferred Qualifications
|
|
||||||
|
Education: |
|
Salary : $110