← All projects

SpeechXP

A Chat-Driven App for speech coaching: users converse with an LLM persona, and structured model outputs grade their delivery in real time on clarity, tone, confidence, and relevance.

Chat-Driven AppPostgreSQLPrismaJWTStable DiffusionNext.js
Screenshot of SpeechXP

Key wins

  • 10,000+ synthetic character images generated with Stable Diffusion
  • Real-time scoring on clarity, tone, confidence, relevance
  • JWT-protected API endpoints, PostgreSQL + Prisma data layer

As a Chat-Driven App

SpeechXP uses the Chat-Driven App pattern to grade real-time dialogue practice. The LLM scores the user’s reply to a simulated high-pressure scenario and emits structured outputs the app turns into a graded dashboard and formatted feedback.

User input

Scenario: explain a missed project deadline to a frustrated client. User reply: "I want to be straight with you — we slipped because of an unexpected dependency change. Here’s the new plan and the date I’ll commit to."

LLM output

Strong, accountable opening. Watch the pace on the recovery plan.

[ACTION:show_grade clarity="8" tone="7" confidence="9" relevance="9"]
[ACTION:format_feedback section="strengths" body="Owned the slip, offered a concrete recovery."]
[ACTION:format_feedback section="growth"    body="Slow down when delivering the new commitment date."]
[ACTION:advance_to next_scenario="2"]

What the app does

SpeechXP parses each tag and renders the score breakdown, lays the feedback into the strengths/growth sections, and queues the next scenario in the practice loop — all driven from the chat surface, not separate UI controls.

Read about the Chat-Driven Apps pattern →

What I built

  • Built an AI speech-coaching product that grades user performance in difficult dialogue scenarios and helps users practice high-pressure conversations.
  • Designed the product around real-time feedback, scoring, repeatable practice loops, and measurable performance improvement rather than passive chatbot interaction.
  • Developed a PostgreSQL-backed application using Prisma for data modeling, persistence, and structured access to user/session data.
  • Implemented JWT-based authentication to protect API endpoints and reduce unauthorized access or endpoint abuse.
  • Applied AI evaluation concepts to assess clarity, tone, confidence, relevance, and effectiveness of user responses in simulated conversations.
  • Used Stable Diffusion to generate 10,000+ synthetic character images for interactive dialogue scenarios and simulated conversation practice.