About Liquid AI

Spun out of MIT CSAIL, we build general-purpose AI systems that run efficiently across deployment targets, from data center accelerators to on-device hardware, ensuring low latency, minimal memory usage, privacy, and reliability. We partner with enterprises across consumer electronics, automotive, life sciences, and financial services. We are scaling rapidly and need exceptional people to help us get there.

The Opportunity

On Liquid’s Applied ML team, you’ll adapt our audio models for enterprise products where a spoken request needs to trigger the right action. When someone asks a voice assistant to “turn down the passenger-side temperature,” the model needs to select the right tool, pass the right arguments, and confirm what changed, even in a noisy cabin.

You’ll work directly with customers to define what reliable behavior looks like, then own the data, post-training, evaluation, and delivery needed to achieve it. What you learn from those deployments will help improve Liquid’s audio models and post-training pipelines.

What We’re Looking For

We need someone who:

Defines success before training. You ask customers for examples of what should happen, what can go wrong, and which mistakes matter most. You turn those answers into tests both teams can use to make a release decision.

Investigates failures hands-on. When a model takes the wrong action, you listen to the audio, inspect the predicted call and training examples, and design an experiment that can distinguish between likely causes.

Makes tradeoffs clear. You explain what the evidence supports, where uncertainty remains, and what you recommend next. If a quality target conflicts with a device’s latency or memory budget, you help the customer choose a workable path.

The Work

Lead the technical work on enterprise audio engagements, from scoping the customer’s use case and target hardware through model delivery.

Build evaluations that combine automated checks with human listening. Measure tool selection, argument accuracy, unnecessary calls, task completion, and response latency across speakers and noise conditions.

Build and curate speech and text datasets from real and synthetic dialogue. Cover requests the model should act on, ambiguous requests that need clarification, and cases where it should take no action.

Fine-tune audio models to translate speech into structured tool calls and natural spoken responses. Choose supervised fine-tuning, preference optimization, or reinforcement learning based on the behavior you need to improve.

Test on the target device, review failures with the customer, and bring reusable datasets, evaluations, and training improvements back into Liquid’s core pipelines.

Experience & Capabilities

What you’ll need

Hands-on experience post-training a language or multimodal model. You can explain the training objective you chose, how the data influenced the result, and how you investigated a regression.

Experience designing task-specific evaluations, breaking results down by failure mode, and using the findings to decide what to change.

Experience building and validating training datasets, including assessing coverage, label quality, and gaps between training examples and real usage.

Familiarity with function calling, tool use, or structured model outputs, including how to check whether a model selected the intended action and supplied valid arguments.

Helpful, not required

Experience fine-tuning or evaluating speech-to-speech, speech recognition, text-to-speech, or multimodal audio-text systems.

Experience delivering an applied ML project directly with a customer.

Hands-on work with preference optimization or reinforcement learning beyond supervised fine-tuning.

Experience evaluating models on devices with tight latency or memory budgets.

What Success Looks Like (Year One)

You’ve independently delivered an enterprise audio model that meets the customer’s agreed quality and latency criteria on the target hardware.

You’ve shipped a voice-driven workflow that selects the correct tool and arguments, asks for clarification when needed, and avoids unintended actions under realistic audio conditions.