shipped field record
AI/ML Data Operations
I worked at TalkingPoints from 2022–2026. Each project lists its own dates, so some begin later within that span.
- Problem
- AI classifiers and translation evaluation needed human review and visible costs.
- My role
- Stewarded classification workflows from experiment to production operations.
- Shipped
- Evaluation apps, sampling loops, and cost safeguards.
- Proof · 2026-07
- Made quality, cost, and tradeoffs reviewable before automation reached users.
Public systems · Agent systems
Message classification stewarded to production. Machine Translation Quality Estimation app for self-service evaluation. Cost safeguards and visible tradeoffs.
Materials & technology
Links
Frequently Asked Questions
How do you evaluate AI classification accuracy when ground truth is expensive?
Sample-based evaluation with human labeling. You can't label everything, but you can label enough to estimate accuracy and identify systematic errors. The key is choosing samples that are representative, not just convenient.
How do you balance AI capability with cost constraints?
Make the tradeoffs visible. Every AI feature has an accuracy level, a latency, and a cost per unit. Build dashboards that show all three. Let stakeholders decide what tradeoff they want, instead of making that decision invisibly in the engineering layer.