Amazon Technologies, Inc.

Michael Dillon

Paul Aksenti Savastinuk

Gopi Prashanth Gopal

Roman Talyansky

Michael Lee Sandige

Eli Osherovich

Charles Winn Degranville

Paul Eugene Munger

Ross David Roessler

Eduard Oks

Amit Dekate

Steven Scott Noble

Joseph James Greene

Sharon Alpert

Steven Xiang Chen

Sepehr Valipour

Christina Alexandra Polyukh

Scott Brown

Matthew Scott Wilding

Systems and methods for embedding-based LLM tuning include generating a first embedding of received user input data and utilizing a translation model trained to associate user input with device names to generate a second embedding that differs at least in part from the first embedding. Reference embeddings corresponding to devices associated with user account data may be generated and a subset of the reference embeddings that satisfy a threshold similarity to the second embedding may be determined. A large language model (LLM) configured to determine a response to the user input data may utilize data for a subset of devices that correspond to the subset of the reference embeddings for performing speech processing.

Embedding-based large language model tuning

This disclosure describes systems and techniques for detecting events, determining a result of each respective event using a first hypothesis source, and calculating a likelihood that a second (and/or additional) hypothesis source would determine the same result of the respective event. The calculated likelihood may then be used to be determine whether to request that the second hypothesis source determine the result of the event, determine an amount of resources of the second hypothesis source to use to make this determination, and/or like.

Generating event results using auto-generated sensor data

Systems and techniques for detecting events involving a user that occurred during a session in a facility, and performing a holistic, session-based analysis of the events to determine items taken by the user from the facility. An inventory management system may initially analyze the events to determine a hypothesis for a result of each respective event and a probability that the hypotheses are accurate. An initial listing of the items determined to be taken by the user may be generated along with probabilities that the items were taken. Upon determining that a probability an item was taken is lower than a threshold, the management system may select an algorithm to process an event in the session that is likely to resolve uncertainty in the initial listing. The management system may iteratively process events until a final listing represents the items taken by the user with an allowable probability.

Session-based analysis of events

Systems and techniques for detecting events involving users that occurred at an inventory location, and then using the events to determines states of the inventory location at various times. For instance, a system may generate event data representing first events that occur within an inventory location. The first events may include an associate performing a task, a customer removing an item, a customer returning an item, and/or so forth. The system may then use the event data to determine a state of the inventory location during a second event. In some examples, the state may include a tidy state or an untidy state. If the state includes a tidy state, then the system may generate a listing of items that includes at least one item associated with the second event. However, if the state includes an untidy state, then the system may further process the second event.

Analysis of events for state detection

The hand which a user is using to hold an electronic device can be determined by analyzing data captured by one or more motion sensors on the device. The curvature to the motion can be indicative of handedness, and processing motion features using a classifier algorithm can enable the determination of handedness with a corresponding confidence. In some embodiments, motion data is collected over a monitoring window, and handedness values are accepted when the handedness value remains the same with at least a minimum confidence for at least a minimum number of window periods. A determination of handedness enables an operating system and/or applications executing on the device to adjust one or more operational or interface aspects in order to make it easier for the user to operate the device using the hand currently holding the device.

Handedness determinations for electronic devices

A device with a movement-based user interface (UI) may establish thresholds that govern when the UI moves and when the UI stays still to avoid interruptions to the user experience based on undesired movement, such as movement caused by a vehicle, etc. The thresholds may be based on detecting a movement type experienced by a device (e.g., bicycle, car, etc.). When movement is detected beyond the threshold the UI may be configured to smooth a displayed UI transition from a first position to a second position to improve the user experience.

User interface configuration to avoid undesired movement effects

A device that operates a user interface (UI) that depends on the position of the user's head. The device estimates the head position using motion sensors. The device establishes an initial estimated position of the head and tracks the device's movement from an initial reference position to determine the new position of the head. If the head crosses a position threshold the device may partially re-center the head to avoid large errors.

Growing community of inventors

Seattle, WA, United States of America

Michael Dillon