Job description
About
We are reimagining the phone. AI phones need to be proactive, intuitive, intelligent and low cost. This is a mission critical for rest of the world to adopt intelligence.
There are over 5 billion people using everyday applications such as email, messaging, notes, tasks, and calendars, yet most are not AI-native. Our mission is to build proactive applications and devices that bring intelligence into daily conversations, coordination and workflows, with minimal prompting.
Responsibilities
Evaluate existing foundation models (e.g. Qwen, Llama, Gemma and other suitable open-weight models) against our target phone use cases.
Identify which capabilities can be handled by existing models and where additional model adaptation or fine-tuning is needed.
Experiment with SFT, LoRA/QLoRA and other lightweight fine-tuning techniques for specific use cases.
Improve model performance for intent recognition, instruction following, tool calling, structured outputs, and multi-step task execution.
Explore synthetic data generation and dataset construction for phone-specific workflows.
Build evaluation datasets and benchmarks to measure model quality, reliability, and task completion.
Compare different models and approaches, including local versus cloud inference.
Work closely with the Agent Runtime team to improve model performance in actual agent workflows.
Requirements
Strong background in Machine Learning, Deep Learning, or NLP.
Practical experience working with LLMs and open-weight models.
Experience with model fine-tuning, such as SFT, LoRA/QLoRA, and preferably RL or preference optimization.
Experience with PyTorch and common ML tooling.
Understanding of LLM evaluation, benchmarking, and dataset preparation.
Ability to run experiments, analyze model failures, and translate findings into measurable improvements.
Nice to have
Experience with agentic systems, function calling, and tool-use training.
Experience with synthetic data generation and preference datasets.
Experience with model distillation and smaller language models.
Experience with deploying models in production environments.
Originally posted on Himalayas
Who can apply
Eligible countries: United States. Accepted UTC offsets: UTC-10, UTC-9, UTC-8, UTC-7, UTC-6, UTC-5, UTC+14. Review the full description for employer-specific work authorization, residency and schedule requirements.