Real-time AI coaching platform
Codexia Live
A real-time coding coach that watches friction signals and responds with progressive hints instead of handing over answers.
Outcome
Converted an open-ended AI tutor idea into a low-latency coaching loop with measurable intervention points.
Problem
Most coding tutors either wait for a prompt or reveal too much. The product needed to detect when a learner was stuck and intervene with the smallest useful hint.
Approach
Modeled the session as an event stream, separated signal detection from response generation, and used staged prompts for nudge, hint, and explanation modes.
Architecture
SvelteKit UI streams learner events to a Cloud Run service over WebSockets. The orchestration layer scores friction signals, retrieves relevant exercise context, calls tools when needed, and emits coaching responses back to the UI.
Result
A credible MVP architecture for live AI coaching, with the system behavior visible enough to tune safely.
Lessons learned
Agentic products need observable policy decisions. Without a readable event timeline, prompt quality becomes guesswork.
Constraints
- Keep perceived latency low enough for a live learning session.
- Handle partial context, silence, and typing pauses without over-triggering.
- Keep the workflow explainable to product and curriculum teams.
Technical decisions
- • Used WebSockets for bidirectional session state instead of polling.
- • Kept intervention policy outside the UI so product rules could evolve independently.
- • Designed prompts around progressive disclosure to reduce answer leakage.
Key features
- • Friction-signal detection from typing pauses and silence.
- • Progressive hints with escalation paths.
- • Session timeline for debugging agent behavior.