Insights from the front-lines of applied AI research for agentic CX
Ada pairs customer-authored intents, full-transcript reading, and a self-improving loop vs. an LLM judge—reaching 90%+ topic and 80%+ intent accuracy.
ReadSolve hard problems. Ship real things.
Join a team building at the intersection of cutting-edge AI and the industry's largest customer conversation dataset.
Open roles
Ada's modular reasoning architecture was replaced with dual reasoning with a re-base eval harness and Legitimacy Classifier. Adversarial pass rate: 88% to 97%.
Ada's AI agents detect and suppress automated emails, preventing infinite loops (~8,000 per week) and enabling earlier, configurable survey delivery.
How Ada built a per-conversation quality-scoring judge that scales to millions of conversations and stays calibrated to each customer's definition of "good."
DE replays production conversations through modified prompts & models, using an LLM-as-judge. Verdicts aggregate into win-rate metrics; traces into themes.
Replacing Ada's fastText language detector with an LLM tool call lifted recall from 78.6% to 97.9%: call the tool first and condition on script, not confidence.
Migrated knowledge and coaching retrieval to turbopuffer and Cohere Embed v4, improving latency, recall, and headroom to scale per-tenant namespaces.