Latitude / Level AI
I connected task-specific evaluation with the requirements of running AI in production.
Speech / Brand-name recognition
Measured · task-specific quality
Accuracy recognizing brand names in speech. This is a task-specific result, not overall model accuracy.
Illustrative evaluation framework · no live model calls
The problem
An overall model score can hide a failure in a detail the customer needs. Brand-name recognition needed its own quality check, alongside the requirements of running AI in production.
My part in the change
For speech recognition, we measured whether transcripts preserved brand names correctly.
My product work connected task-specific quality with the operating requirements of customer-facing products.
I worked on production operating costs alongside the quality requirements of AI capabilities.
Scope & contribution
Latitude is Level AI’s shared AI layer for customer-experience products. My product work connected task-specific evaluation with quality, latency, reliability, and cost. This case focuses on recognizing brand names in speech and the operating requirements behind the product.
An overall model score can hide failures in details a customer cares about. A transcript needs to preserve the right brand name; a summary needs to retain important context. Each task needs its own quality checks, alongside response time, reliability, privacy, and cost.
I’ve led product work on task-specific model evaluation and the cost of running AI capabilities in production. This case focuses on speech-model brand recognition and operating costs within the broader Latitude platform.
Decision context
A brand name matters even when the overall transcript looks accurate.
For speech recognition, we measured whether transcripts preserved brand names correctly, rather than relying only on a broad model score.
The cost of running an AI capability is part of its product requirements.
My work connected task-specific quality requirements with response time, reliability, and operating cost.
Published product walkthrough
Published product imagery illustrates behavior, not feature-release dates during my tenure. Interactive frameworks are explicitly illustrative.
Speech / brand recognition
Speech-model brand-recognition accuracy.
A task-specific reported outcome, not overall model accuracy.Illustrative evaluation framework · no live model calls
My contribution
Transcription, redaction, intent, summaries, inferred CSAT, quality assurance, and Voice of Customer solve different problems.
My product work has covered all seven capabilities: transcription, redaction, intent, summaries, inferred CSAT, quality assurance, and Voice of Customer. That work included task-specific evaluation, such as whether speech models correctly recognize brand names.
Evaluation needs to reflect the task and the consequences of getting it wrong.
Model capability needs evaluation and feedback within the product.
Outcomes in context
Recognition of brand names in speech improved from 71% to 95%. My broader product work balanced task-specific quality with latency, reliability, and operating cost.
Technical product judgment means connecting model behavior to the customer experience—and to the economics of running it.
Visit the product at Level AIKeep exploring
AI platform + product organization
GOOD PRODUCTS START WITH A CONVERSATION.