Dolores

Redefining business data verification through a voice bot

Voice/chat bot

Designer & bot trainer

Timeline
2021
‍
Team
CTO, 3 Engineers, and myself

Highlights
‍Used by dataPlor's validations team, handles +100k calls a day on several countries to verify POI’s at a fraction of the cost of a manual review.

Main screen of Kuebiko's data manipulation tool with SQL tool

Lo-fi of the Dolores validation platform

Overview

Dolores was an internal tool built to automate business data verification through a custom voice call bot. Designed before the major breakthroughs in modern TTS and STT, the system had to work around technical limitations while supporting thousands of live calls daily.

Alexandra VV · Dolores Question

The Problem

How can we automate voice-based data collection and validation in a way that feels natural, captures structured output, and is resilient to failure, using tech that was still evolving?

Components of this call bot

Approach & Research

We interviewed analysts, and operations leads to understand how verification decisions were made. We also tested existing speech-to-text APIs and documented their failure points. Because the calls were made to LATAM countries, we had to evaluate multiple telephony providers. I set up a mock call center to benchmark audio quality, latency, and speech recognition accuracy under real-world conditions.

This helped us identify the tools best suited for regional accents and variable connectivity, even though the tech was still far from ideal. Most calls broke down not due to intent, but phrasing, accents, colloquialisms, or slight deviations from expected answers. This informed our fallback design and led to modular prompts that could recover gracefully.

SSML example

Pain points

Manual validation didn't scale

Human agents could only handle a limited number of verifications per day, causing delays and inconsistent quality.

Off-the-shelf tools weren't adaptable

At the time, available voice tech lacked the flexibility to script domain-specific interactions or handle fallback logic gracefully.

No UI for data validation

Debugging, and validating the data required engineers. There was no shared space for non-technical team members to improve the experience.

What I Did

Mapped typical call scenarios and edge cases with operations
Worked closely with engineering to define and test how prompts, intents, and fallback logic were triggered across varied inputs
Designed an interface where internal teams could review, rate, and annotate call transcripts for quality and intent coverage
Created documentation for tone, phrasing, and escalation patterns to ensure consistency across the system

UI of Bot's Auditing Tool

Key Work

A hypothesis

Design a voice system that works even when the tech is not quite there.

We hypothesized that clarity, repetition, and modular fallbacks would outperform complex NLP in high-volume, low-attention environments. Our focus became consistency over cleverness.

Call Review Interface
Designed a tool for internal teams to validate chatbot behavior through transcript review, feedback input, and quality scoring.

Fallback logic
Designed layered fallback responses that adapted based on silence, confusion, or contradiction, aiming to minimize drop-offs.

Tone and Timing
Refined pacing, pauses, and language to reduce misfires from overlapping speech or slow responses.

‍Language Adaptation
Studied colloquialisms and regional phrasing common in LATAM to reduce friction and improve bot understanding across varied speech patterns.

Bot Auditing tool's different iterations

Outcomes

Reflection

Dolores pushed the limits of what was possible at the time. It taught me to design for failure, plan around imperfect systems, and always keep the human fallback in mind.

Today, tools like OpenAI Whisper or ElevenLabs would make Dolores faster and easier to build, but back then, every step had to be engineered.

Boosted data validation accuracy by 60%.
Performed 100K+ daily calls in Brazil and Mexico.
Improved validation speed of the calls.