AI Modeling Engineer

Palona’s AI agents operate in real restaurant environments: noisy phone lines, varied accents, complex menus, interruptions, incomplete information, strict business rules, and customers who expect an immediate, natural response. Improving these systems requires more than selecting the newest model. It requires disciplined evaluation, high-quality data, modeling judgment, experimentation, and production feedback loops.

We are looking for an applied AI Modeling Engineer to improve the intelligence, accuracy, safety, latency, and cost of Palona’s voice and multimodal agents. You will own problems across model selection and routing, prompting and context, fine-tuning or post-training when justified, speech and language quality, evaluation methodology, dataset development, and model behavior in production.

This is a product-facing modeling role. Research depth matters, but success is measured by improvements that survive contact with production and create better guest, restaurant, and business outcomes. You will work closely with product, full-stack, infrastructure, and customer-facing engineers to move from hypothesis to experiment to reliable deployment.

• Develop modeling and experimentation strategies for high-impact agent problems in voice, language, reasoning, ordering, multilingual behavior, and multimodal understanding.
• Build rigorous offline and online evaluations that measure task completion, accuracy, safety, latency, cost, conversational quality, and business outcomes.
• Create and maintain representative datasets from simulations, human annotation, production feedback, and difficult edge cases while protecting sensitive data.
• Evaluate frontier and open-source models and make clear build, buy, route, prompt, fine-tune, or distill decisions.
• Improve prompting, context construction, memory, tool-use policies, structured outputs, model routing, and fallback behavior.
• Design fine-tuning, preference optimization, distillation, or other post-training work when it offers a measurable advantage over simpler methods.
• Partner with speech and real-time engineers to improve ASR, TTS, turn-taking, interruption handling, pronunciation, multilingual behavior, and end-to-end latency.
• Develop analysis tools that explain model failures, slice performance by scenario, detect regressions, and accelerate iteration.
• Ship model changes with production guardrails, staged rollouts, monitoring, rollback paths, and clear quality gates.
• Translate new research and model releases into concrete product opportunities and communicate tradeoffs to technical and non-technical partners.
• Raise scientific and engineering standards through reproducible experiments, thoughtful reviews, and clear documentation.

• 3+ years of industrial experience in relevant technical domain.
• Strong machine learning foundations and hands-on experience developing or evaluating production AI systems.
• Strong Python skills and experience with modern ML tooling such as PyTorch, JAX, Hugging Face, or equivalent systems.
• Practical experience with LLMs, speech models, multimodal models, or agentic systems.
• Ability to design reliable experiments, define useful metrics, analyze noisy results, and avoid optimizing against weak proxies.
• Experience building datasets, evaluation harnesses, model services, or training and inference pipelines.
• Strong software engineering judgment; your work is reproducible, tested, observable, and usable by other engineers.
• Ability to connect modeling choices to product constraints including latency, cost, privacy, safety, and user experience.
• Comfort operating in ambiguity and collaborating across research, engineering, product, and customer contexts.
• AI-native working habits and genuine curiosity about new model capabilities and limitations.

• Competitive Salary and Stock Option Plan.
• Medical, dental, vision, retirement, leave, and disability benefits as applicable.
• Family Leave
• Short Term & Long Term Disability
• Paid time off and company holidays.
• Learning and development support.