Contents

Odaptos' Path to Scalable AI-Driven User Testing

Odaptos is transforming the way companies understand their users. Their innovative platform goes beyond traditional user testing — it incorporates AI-based emotional analysis. Odaptos captures and interprets users’ emotions during interviews, giving businesses a much clearer idea of the user experience.

The platform’s smart video player breaks down the videos into key moments, highlighting where users face difficulties and, most importantly, how they feel during these moments. This combination of functional and emotional insight gives companies a clearer picture of what works and what doesn’t in their products or services.

Jump straight to a key chapter

Project Description
Work Agenda
Tasks and Challenges

Results
DevOps Tech Stack

Project Description

Odaptos came to IT Outposts with an infrastructure primarily built on Lambda functions. Their entire system was made up of scripts running different tasks. However, as the company aimed to scale and integrate new services, they recognized the need for a more robust and flexible infrastructure. The client’s objectives were to fine-tune their existing setup.

Work Agenda

Client

Odaptos

Location

France

Technical team

1 DevOps engineer
1 SRE engineer
Project manager

Project timeframe

August 2024 - February 2025

Project goals

Resolve intermittent 503 errors to improve reliability and user experience.

Develop a queue-based processing architecture capable of handling multiple concurrent requests.

Transition Lambda-centric services to a more flexible, container-based solution using Amazon ECS.

Optimize operational costs through replacing multiple third-party translation APIs with our own speech-to-text service and introducing scheduled service operations.

Implement a cost-effective, dynamic GPU-enabled processing system for AI model integration.

Enhance data security by implementing client-specific encryption for uploaded content using KMS.

Tasks and Challenges

Resolving intermittent 503 errors

Odaptos was experiencing periodic 503 errors that were causing significant downtime. Our team quickly identified the root cause of these errors and resolved the issue within a day. This made the platform way more reliable and better for users.

Transitioning from Lambda-centric architecture

The existing Lambda-based infrastructure had limitations in terms of scalability and flexibility, particularly for running GPU-intensive tasks. As a solution, our team released two microservices on container-based Amazon ECS.

STT handles speech-to-text functionality. Named Entity Recognition (NER) is the second service, which processes transcripts and extracts keywords, and we developed it from scratch. Both services were deployed in containers, which allowed us to overcome the previous infrastructure limitations.

Implementing GPU-enabled processing

Odaptos aimed to integrate AI models that require GPU capabilities, but the expense of continually operating GPU-enabled instances is too high. To address this issue, our engineers created a system that automatically activates GPU-enabled machines as needed, runs the required tasks, and then powers them down. This greatly lowers costs while providing the essential processing power.

Configuring a queue system

Initially, Odaptos could only process one request at a time. For example, if five users simultaneously submitted audio processing requests, only one would be processed while the others would fail.

To address this issue, we redesigned the codebase for both the STT and NER services to work with queue systems. Ultimately, all incoming requests can be automatically placed in a queue for orderly processing, which now allows the system to handle multiple concurrent requests.

Optimizing operational cost

Another challenge Odaptos faced was their reliance on multiple cloud providers for various services. They used certain services from Amazon, others from Azure, and yet others from Google.

This multi-cloud approach was necessitated by specific capabilities offered by different providers. For instance, Azure provides French language translation, a feature not available in Amazon's offerings. However, this not only increased the complexity of infrastructure but also potentially raised costs.

Given that we switched to the STT service, which uses Whisper technology, we no longer required the cloud’s APIs mentioned above because our STT service could handle all of the tasks.

We also set up the STT and NER services to run on a schedule — eight hours per day — to save money. Our team made this work efficiently by using queue systems, which allowed more users to use the services during the limited operating hours.

Enhancing security measures

Odaptos places a high priority on data security and wants to implement client-specific encryption for video uploads. We configured a system using Key Management Service (KMS) to generate unique encryption keys for each client.

Contacts

Are you developing an AI-based project and seeking the ideal infrastructure solution? IT Outposts knows how to tackle the challenges of scaling AI models. Reach out to us today, and let’s architect the future of your AI project together. Your next breakthrough is waiting!

Discuss your project

Results

Our collaboration with Odaptos has yielded positive results. We’ve addressed their urgent infrastructure challenges, and the improvements are evident: their services are now operating more smoothly.
Plus, thanks to deploying the STT service, configuring a queue system, and setting up scheduled service performance, our team managed to significantly lower the operational cost for our client.