BTH works in education and wanted to give personalized support to more than 30,000 students. People alone could never cover that. Instead of building a one-off app, we designed an AI product with them: a core architecture of conversational AI agents with avatars that speak, listen and respond in real time.

On that same foundation we launched three different products. This case shows what it means to productize artificial intelligence: invest once in the right architecture, then turn it into several solutions that create value.

BTH's challenge had four fronts:

• Scale: more than 30,000 students who needed live sessions and personalized attention.
• Specific knowledge: answers had to come from each course's content, not from generic information on the internet.
• Cost: AI with voice, avatars and frontier models can be expensive per session, and the model had to stay sustainable at scale.
• Several needs, one investment: tutoring, entrepreneurship and career guidance. Building three separate systems would have tripled the time and cost.

We didn't start with the app. We started with the platform. We designed a core architecture that separates what is shared (voice, avatar, agents, infrastructure) from what changes in each product (knowledge, staging and criteria). Its pillars:

• Multi-tenant: each institution, program or course has its own space, configuration and data on the same infrastructure.
• Multi-agent: each agent has its own RAG, fed with the materials for its course, so it answers precisely about what the student is learning.
• Natural conversation with an avatar: speech-to-text (STT), reasoning with frontier models, text-to-speech (TTS) and lip-sync on Ready Player Me 3D avatars. Students talk with a tutor who listens and answers while looking at them.
• Serverless: the platform scales automatically with demand and only uses resources when sessions are active.
• Highly customizable: avatar, voice, personality, knowledge and evaluation criteria are configured without rebuilding the system.

1. AI avatar tutors for live sessions

The first version is a multi-tenant, multi-agent platform that serves more than 30,000 students in live sessions with conversational AI agents. Each agent specializes in one course and answers students' questions before their final exams, with a patience and availability no human team can offer at that scale.

2. Georgia: an AI elevator pitch evaluator

On the same architecture we built Georgia, a specialist in evaluating elevator pitches to promote entrepreneurship in education. We designed the staging and all the data processing so each student presents a pitch and receives an evaluation based on fixed criteria. That way, each of the 30,000 users gets the same objective evaluation and concrete feedback to improve.

3. AI career guidance

The third version helps students decide their future. It combines recognized vocational psychology tests, which identify the careers that best fit each profile, with up-to-date information from the web and employment databases.

The result is a personalized report with recommended careers, universities and their costs, expected salaries, funding sources and other opportunities, with real names and links for universities and companies. To get there, we combined structured data (employment databases) with unstructured data gathered through web scraping, so the information stays current and relevant for each student.

1. Real-time conversation: chaining speech-to-text, a language model, text-to-speech and lip-sync with the flow of a natural conversation.
2. Cost per session: using frontier models without costs spiking across thousands of students.
3. Accuracy: avoiding made-up answers and keeping each agent faithful to its course content.
4. Consistent evaluation: making sure Georgia scored every pitch on the same criteria, regardless of the student or the moment.
5. Reliable web information: turning scattered, changing data into a clear, verifiable report.

• Reusable architecture: a shared core for voice, avatar, agents and infrastructure, plus a configurable layer for each product.
• RAG per agent: each agent queries only its own course knowledge, which reduces hallucinations.
• Aggressive cost optimization: model selection, context management and serverless infrastructure to reach about COP $500 (Colombian pesos) per session with frontier models.
• Criteria-based evaluation: fixed rubrics and structured processing of each pitch so Georgia is objective and repeatable.
• Combining structured and unstructured data: web scraping, employment databases and source validation to deliver up-to-date reports with real links.

• 3 products on 1 core architecture: the second and third products reused the foundation instead of starting from scratch.
• More than 30,000 students served in live sessions with AI agents.
• About COP $500 (Colombian pesos) per session using frontier models, a cost that makes conversational AI viable at scale.
• Scalable and customizable: new institutions, courses and products are added through configuration.

The BTH case sums up our method:

1. Design the foundation, not the app: plan from day one for what will be reused.
2. Separate the shared from the specific: voice, avatar and agents are shared, while knowledge, staging and criteria are configured.
3. Control unit cost from day one: an AI product only scales if every session is profitable.
4. Launch new versions on what you've built: each new product costs less and ships faster than the last.

This is the same approach we apply in our Custom Software and AI Agents, Data for AI and AI Agent Audit and Evaluation services: build it once, build it right, and multiply the value.