Whisper through Groq for transcription
Server-side transcription gives consistent accuracy in Spanish and lets the evaluation run asynchronously after the recording.
Instead of: The browser's Web Speech API
Building an AI public speaking and debate trainer with FastAPI, React, Whisper and Llama 3
a.k.a. My Thesis→ CONTEXT: Undergraduate thesis in Systems Engineering, Universidad del Valle (2025), built on the soft-skills platform started by Juan Azcárate
→ OFFICIAL TITLE: Creación de módulos interactivos para el desarrollo continuo de la oratoria y el pensamiento crítico en el prototipo de enseñanza de habilidades blandas
ROLE: Full Stack Developer · Researcher
TIMELINE: Feb 2025 – Jun 2025
STATUS: Live
FOR: Universidad del Valle · Systems Engineering undergraduate thesis
STACK:
frontend
- React 18
- MUI
- react-webcam
backend
- FastAPI
- SQLModel
- PostgreSQL
ai
- Whisper (Groq)
- Llama 3.3 70B (Groq)
- GPT-4o mini (OpenAI)
infra
- Vercel
- Auth0
TEAM: 2 people · with Juan Azcárate (Base platform author), Joshua David Triana Madrid (Thesis advisor)
BLOCK // PROBLEM
Systems engineering students needed a practical way to improve communication and critical thinking skills to strengthen their academic and professional profile.
BLOCK // SOLUTION
A FastAPI and React platform with modules for oratory, critical thinking and debate. Speech exercises are recorded in the browser, transcribed with Whisper through Groq and evaluated asynchronously; Groq's Llama 3.3 70B and OpenAI's GPT-4o mini generate the debate topics and structured, personalized feedback. PostgreSQL stores progress and Auth0 handles sign-in.
BLOCK // CONSTRAINTS
- One academic semester for design, build and user testing
- An inherited codebase (React 18, MUI, Auth0) that had to keep working for its other modules
- Spanish-speaking students, so transcription and feedback had to work well in Spanish
ATTACHED EVIDENCE
ENGINEERING IMPACT REPORT
An end-to-end AI learning platform that achieved 95% usability approval and measurably improved both oratory confidence and argument coherence in systems engineering students.
85%
Students improved speaking confidenceSelf-reported after using the oratory module92%
Participants with stronger debate coherenceMeasured in debate responses after the module95%
Positive usability feedbackUsability tests with participating studentsENGINEERING DECISIONS
Two LLM providers: Llama 3.3 70B on Groq and GPT-4o mini
Groq covers the volume at speed and no cost; GPT-4o mini is a second provider for topic generation and feedback when quality or availability calls for it.
Instead of: A single provider
Asynchronous evaluation instead of a blocking request
Transcription plus LLM feedback takes seconds. Running it in the background keeps the UI responsive and lets students move on and come back.
Instead of: Waiting on the request while the model answers
Build on the existing soft-skills platform
Reusing its auth, dashboard and database meant the semester went into the oratory, critical-thinking and debate modules instead of scaffolding.
Instead of: A greenfield application
FIELD NOTES & DEVELOPMENT HIGHLIGHTS
NOTE_01
Two complete learning modules for oratory and critical thinking, from UX to AI integration
NOTE_02
Browser recording, Whisper transcription and LLM evaluation wired into an asynchronous feedback pipeline
NOTE_03
Usability tests with systems engineering students that shaped the final version of both modules
LESSONS LEARNED
- Transcription quality drives the whole feedback loop: a weak transcript produces confident but wrong feedback.
- Asynchronous evaluation changed the experience: students could record, move on and come back to their feedback.
