Technology
We build real-time AI-assisted systems that amplify teachers. The core of Oktatron is a structured feedback loop between students, AI, and human educators.
Real-time AI-assisted lesson loop
1. Student input
Students solve exercises via tablet, whiteboard, or camera. Input is processed in real-time.
2. AI processing
LLM-powered generation and evaluation. Adaptive task creation based on detected weaknesses.
3. Teacher control
The teacher remains in control. AI assists, highlights patterns, and reduces workload.
4. Continuous adaptation
Feedback loops improve personalization and long-term development.
Core system components
We design for real-time sessions, low-latency feedback, and safe iteration of AI features in production.
Real-time sessions
Video + audio + collaboration designed for 1 teacher and multiple students, with teacher-first control.
LiveKitWebRTC
Event-driven backend
Services communicate through streams for scalable processing (task generation, evaluation, analytics).
Go (Fiber)NestJSGraphQLRedis Streams
Data + personalization
We store structured learning state and analytics to personalize practice and track progression.
PostgreSQLObject storage
LLM + evaluation
We ship AI features with evaluation harnesses and regression checks — correctness matters in education.
PythonLLM toolingTest sets
Cloud infrastructure
Built on GCP with containerized deployments, designed to scale from Hungary to the US.
GCPCloud RunGCSDocker
Reliability
Observability, sane deploys, and operational excellence — reliability is a first-class feature.
MonitoringLoggingCI
System flow
A simplified view of how real-time lessons, AI services, and data storage connect.
Teacher + Students
Live session (video/audio/whiteboard)
→
API Gateway
Auth · GraphQL · orchestration
→
Stream Processing
Generation · evaluation · analytics
LLM Services
Prompting · tools · eval harness
→
Learning State
Student model · progress · weaknesses
→
Feedback to lesson
Adaptive tasks + teacher insights
Why this is technically hard
Low latency + AI
Real-time classroom interaction cannot wait seconds for AI responses. We must balance latency, cost, and model complexity.
Correctness matters
In mathematics education, hallucinations are unacceptable. Evaluation, regression testing, and verification are essential.
Hungarian language support
Most LLM tooling is English-first. Educational AI in Hungarian requires careful prompting and evaluation.
Human-in-the-loop design
Teachers must remain in control. AI assists but never replaces professional judgment.
Scaling real-time sessions
Video, streams, AI processing, and storage must scale from Hungary to US-level concurrency.
Evaluation & safety
Shipping AI in education requires strong internal evaluation, safety checks, and iteration discipline.
Research mindset, production discipline
Oktatron is built from real teaching experience and research-driven engineering. We prototype fast, but we ship with discipline: evaluation harnesses, regression checks, and clear ownership.
Teacher-first design
Human presence stays central. AI reduces background workload.
Measure what matters
We care about learning outcomes, not just model outputs.
Scale the impact
Personalized education should be available to far more people.
Explore open rolesWe’re hiring builders across platform, AI/LLM, and reliability.