Member of Technical Staff, Pretraining Evaluations

Hark is an artificial intelligence company building advanced, personalized intelligence. One that is proactive, multimodal, and capable of interacting with the world through speech, text, vision, and persistent memory.

We're pairing that intelligence with next-generation hardware to create a universal interface between humans and machines. While today's AI largely operates through chat boxes and decade-old devices, Hark is focused on what comes next: agentic systems that interact naturally with people and the real world.

To get there, we're developing multimodal models and next-generation AI hardware together - designed from the ground up as a single, unified interface for a new era of intelligent systems.

The Omni team at Hark is building the next generation of AI experiences beyond text, enabling models to understand and generate content across multiple modalities, including text, audio, and beyond. Our goal is to create seamless, real-time multimodal intelligence that powers intuitive and immersive user experiences.

As part of the Omni team, you will build the multimodal foundation models at the core of this effort. You'll build the systems used to evaluate Hark's models during pretraining and mid-training. You'll track held-out loss, measure capabilities across text and multimodal tasks, and make comparisons between training runs reproducible. Your analysis will help researchers choose training recipes, select checkpoints, and identify regressions.

• →Build and maintain an evaluation harness for held-out loss, performance across data domains, and capabilities in text, code, math, and reasoning.
• →Design evaluations for early checkpoints and smaller models using log-likelihood scoring, few-shot prompts, and generated completions.
• →Assess benchmark coverage and saturation, quantify uncertainty in results, and work with data researchers to detect training-data contamination.
• →Version datasets, prompts, graders, and inference settings so results can be reproduced and compared across checkpoints and training runs.
• →Analyze experiments and work with researchers to identify useful recipe changes, select checkpoints, and prioritize capability improvements.
• →Develop new benchmarks for capabilities that existing evaluations do not adequately measure.
• →Extend evaluation tools to image, video, speech, and audio tasks in collaboration with multimodal researchers.

• Experience building model evaluation systems or experimental tools for reproducible comparisons across models or checkpoints.
• Knowledge of language model evaluation methods, including log-likelihood and generative scoring, prompt sensitivity, few-shot variance, and formatting effects.
• Strong statistical judgment, including uncertainty estimation, sample-size planning, and distinguishing statistical significance from practical improvements.
• Experience identifying benchmark contamination, saturation, or gaps between benchmark scores and the intended capability.
• Strong Python engineering skills, including building reliable evaluation pipelines that run efficiently across many checkpoints.
• A record of developing useful benchmarks, evaluation methods, or tools that inform research and engineering decisions.

• Experience designing benchmarks with human ratings or model-based graders.
• Experience studying capability scaling or predicting larger-model performance from smaller experiments.
• Experience with code execution sandboxes, math verifiers, or other programmatic graders.
• Experience evaluating checkpoints during pretraining or mid-training.
• Multimodal evaluation experience: audio, speech, vision, or interleaved inputs.
• Experience evaluating model quality under constrained inference budgets.