Technical Product Manager, Trust & Safety

Innodata Inc.

22h•$180,000 - $220,000

About The Position

As a Technical Product Manager for Trust & Safety, you will define and ship data products that frontier labs use to train and evaluate models for personalization, computer-use, and GDPval tasks safely and trustworthily. You will own the full product lifecycle: translating research needs into specifications, designing reinforcement learning environments, rubrics, and evaluations, coordinating delivery with expert workforces, and measuring whether the data improves model performance. A central part of the role is building sandboxed RL environments with deliberate traps to expose unsafe or untrustworthy model behavior. These include misalignment traps that test for reward hacking, deception, sandbagging, oversight evasion, and power-seeking; cyber traps such as planted credentials, injected instructions, and poisoned tools; and other honeypots that reveal behavior under reduced perceived oversight. You will define safe-behavior rubrics, verifiers, and reward signals; establish measurable safety and trustworthiness metrics; identify blind spots; prioritize product work; collaborate with research, engineering, red-teaming, policy, and delivery teams; monitor emerging threats; and communicate specifications, customer briefs, and safety findings to technical and executive audiences. The role calls for comfort with ambiguity, technical depth, and a commitment to making powerful AI safe and beneficial.

Requirements

At least 5 years of product management experience, including technical products in ML, AI data, security, or trust and safety.
Hands-on familiarity with LLM training and evaluation, including SFT, RLHF/RLAIF, reward models, verifiers, and benchmark design.
Working knowledge of agentic AI, including tool use, computer-use agents, multi-step planning, and associated failure modes.
Strong grasp of AI safety concepts such as reward hacking, deceptive alignment, jailbreaks, and prompt injection.
Technical depth sufficient to read code, review environment specifications with engineers, and make sound technical tradeoffs.
Experience designing metrics that capture risk, system performance, and user impact, and using them to make clear decisions.
Track record of shipping 0-to-1 products in fast-moving, ambiguous environments.
Ability to explain complex technical ideas clearly to researchers, customers, and executives.
Creative, adversarial thinking and willingness to consider model behaviors beyond existing checklists.

Nice To Haves

Comfort with Python and sandboxed environments such as Docker, virtual machines, and browser automation.

Responsibilities

Translate frontier-lab research goals into product specifications for safety training data, RL environments, and evaluations.
Design sandboxed, trap-based RL environments that test for misalignment, unsafe cyber behavior, and other risks.
Specify safe-behavior rubrics, verifiers, and reward signals for agentic tasks, including when models should refuse, escalate, or ask for confirmation.
Define measurable safety and trustworthiness evaluations, including attack success rate, over-refusal, trap trigger rate, and task completion under adversarial pressure.
Track evaluation blind spots and drive iteration.
Prioritize product work by framing problems, assessing business and technical tradeoffs, and distinguishing MVP scope from an ideal state.
Partner with lab researchers, ML and platform engineers, red-teamers, policy experts, and delivery leads overseeing expert annotator pools.
Monitor agentic AI risks, alignment research, and emerging attack techniques, and translate them into roadmap items.
Write product specifications, customer briefs, and safety findings for researchers and executives.
Build reusable trap libraries and quality frameworks that enable delivery teams to create environments efficiently while maintaining reliable verifiers and hard-to-detect traps.