Kea AI: ML Ops Engineer
Skills: Machine Learning, LLMs, NLP, TypeScript, Python, SQL
From December 2023 through October 2026 I had the opportunity to work with Kea AI as an ML Ops Engineer. Kea AI is developing a Voice AI for automating restaurant phone ordering, allowing workers to focus on serving customers in house.
While at Kea, I initially worked on managing SNIPS model lifecycles for Intent Recognition and Named Entity Recognition (NER) tasks, and later expanded to include deploying LLM agents for call handling workflows and developing React-based UI components for user data management.
SNIPS ML Ops Pipeline
Kea’s initial product offering focused on training in-house SNIPS models for intent recognition and NER tasks within the call flow. I was responsible for taking brand-specific datasets and fine-tuning our general production models to improve performance on brand-specific menus and utterances. I automated the separation of the general dataset into training, validation, and testing sets, and evaluated model performance using Kolena to iteratively improve datasets and work toward automation of the model deployment lifecycle.
Menu Transcoding and Standardization
Our NER data initially consisted primarily of simple item and option utterances without a defined structure or hierarchy between menu entities. When Kea shifted product directions, this was a core issue in the refactor. For this task, I reviewed the 8 different point-of-sale (POS) systems supported by Kea and established a standardized menu data format that would be able to preserve essential details and relationships between menu entities. I then developed a transcoder to parse each POS system's menu format into this common standard, enabling more efficient menu resolution and validation within the call workflow.
LLM Agent Expansion
I was the primary engineer responsible for managing our LLM prompt templates for the core call handling agent we deployed, making them generalizable across brands at different tiers and routinely updating them as new features were developed. I also contributed to the creation of React-based frontend UI components that allowed brands to customize details, messages, and responses. I then spearheaded the integration of this data into our prompt templates, allowing us to dynamically generate a unique prompt for each individual call and personalize the experience for every caller. Additionally, I worked extensively on the backend order validation and resolution systems, using call analysis feedback in order to improve order accuracy and reliability.