Dolores - Redefining Business Data Verification with a Chat Bot
Role: UX/UI Designer / Project Manager
Year: 2022
Introducing Dolores
"Have you ever questioned the nature of your reality?" These intriguing words kick-start our journey into the world of Dolores, the name of a revolutionary call bot repository. Dolores plays a pivotal role in confirming data for small businesses in LATAM, continually updating vital information. Remarkably, Dolores achieves an impressive success rate of over 60% during calls, all at a fraction of the cost compared to traditional methods like call centers—an alternative I thoroughly explored.
In collaboration with a team of talented developers, I spearheaded the evolution of this project from an MVP to a versatile multi-language bot capable of handling thousands of daily calls. As the product lead, my role immersed me in a wide array of technical domains, including:
- Natural language processing
- Text-to-speech and speech-to-text technologies
- Telephony providers
- SSML (Speech Synthesis Markup Language) and phonetics
- Bots and machine learning
Bot Audio Sample
The Anatomy of a Voice Bot, Plug-and-Play Style
Creating a call bot requires the integration of numerous components. No single off-the-shelf solution is scalable for the high volume of calls we needed to manage. Consequently, we engineered a custom solution tailored precisely to our unique use case. Our system needed to handle thousands of daily calls, extract information for transformation into JSON, support diverse dialog scenarios, and remain adaptable to accommodate additional languages in the future.
My role extended beyond technical expertise. I translated complex technical jargon into comprehensible concepts, ensuring that the entire team could grasp and apply them effectively in their daily tasks and product requests.
Dolores comprises four primary components, a structure common to many call bots:
Numerous cloud service providers offer these components, such as Google Dialogflow and wit.ai, each with excellent NLP platforms. My considerable efforts were dedicated to evaluating the suitability of different options for each component of our system. This evaluation considered factors like processing speed and latency, crucial when interacting with busy business owners who cannot afford even a millisecond of additional processing time.
To determine the optimal combination, I conducted a series of A/B tests, encompassing technical metrics like request processing time and revenue-driven metrics such as the percentage of confirmed business data. These tests gauged the effectiveness of different stack combinations in obtaining valuable business information.
UI of Bot's Auditing Tool
Decoding Colloquialisms and Teaching the Bot the ABCs
At its core, Dolores's mission is to make thousands of daily calls to businesses, verifying their business names and addresses. While the goal is straightforward, the journey is multifaceted. We initiated the project with Spanish as our primary language since it is widely spoken across Latin America. However, the nuances of natural language processing (NLP) varied significantly from one country to another. Every country had its unique telephone etiquette, even for simple responses like "yes," which varied depending on the region and country.
When deploying the bot in different countries, we faced two primary challenges. First, we had to master the art of combining SSML, phonetics, and colloquialisms to fine-tune the bot's tone and language. We established a standardized call structure with expected outcomes, but the nuances in the bot's language and tone required precision. Our goal was to keep each call under 30 seconds while maintaining effectiveness.
.png)
Once we achieved the desired script refinement, the next step involved training the bot to handle a myriad of potential responses. Rather than crafting responses for every conceivable situation, we utilized examples from previous calls to teach the NLP system. Over time, it began to recognize variations in responses, greatly enhancing the bot's adaptability.
Preventing Bot Anarchy
To ensure the bot's quality and accuracy, I designed an auditing tool that enabled human validators to review Dolores's calls. This tool served several crucial functions:
- It provided a detailed overview of each call's results, including successful connections, responses to prompts, and how data was parsed into our database.
- Validators could either confirm or correct the bot's results.
- In cases of errors, validators could pinpoint and rectify them.
.png)
This ongoing auditing process significantly improved our overall performance, maintaining a high-quality threshold. The process was easily analyzable, producing data for querying and generating insightful reports.
Bot Auditing tool's first iterations
This project has granted me invaluable insights into the intricacies of building chat and voice bots, while emphasizing the importance of data-driven solutions. Conversational UX became a prominent facet of my expertise and interests.
I remain passionate about this subject and am committed to continuous learning in this field.In summary, Dolores has redefined data verification for small businesses in LATAM, delivering precision, efficiency, and scalability through innovative voice bot technology. It exemplifies the power of technology and data-driven solutions in empowering businesses and streamlining processes.
Other Projects
.png)
