Building Theoder: Our Technical Journey with LLMs
Engineering Team
12 min read
Building Theoder: A Technical Deep Dive
Building a truly effective AI tutor requires more than just a prompt and an API key.
The Knowledge Graph
At the heart of Theoder is a persistent Knowledge Graph. Most AI applications treat each session as a blank slate. Theoder, however, remembers everything you've mastered and everywhere you've struggled.
We use a combination of vector databases and graph databases to map the relationship between micro-topics in the UPSC syllabus.
Optimizing for Voice
Learning is faster when it's oral. To make voice coaching feel natural, we had to optimize our pipeline for extremely low latency.
- Streaming VAD: Voice Activity Detection that feels human.
- Edge Processing: Moving as much computation as possible close to the user.
- TTFT (Time to First Token): We've optimized our prompt chains to start speaking back to the student in under 800ms.
Conclusion
We're building a new kind of "thinking tool." It's not just about AI; it's about the intersection of cognitive science and engineering.
Want to experience Socratic AI tutoring yourself?
Start your free diagnostic