Present-iq

The only presentation assistant that learns how you speak during rehearsal, then coaches you with Nemotron AI and follows you automatically on presentation day.

Comment

Inspiration

I was once called on to read aloud in class and my voice shook badly. Afterwards I noticed how many people share that fear. PresentIQ won't cure stage fright, but preparation helps, so I built a tool for the preparation.

My first idea was a teleprompter that follows your voice. It worked until I said the same word twice and it jumped to the wrong sentence. I tried several ways of matching speech to the script, and each failed the same way: browser speech recognition makes mistakes, and speakers repeat themselves. So I changed the question. Instead of chasing a voice live, learn the speaker beforehand.

What it does

PresentIQ is a rehearse-first presentation tool:

  • Script Doctor. Paste a draft and Nemotron flags hedging words, sentences too long for one breath, and the one change worth making first.
  • Rehearse. The script scrolls at a speed you set while the app tracks pace, filler words and a live score.
  • Debrief. When you stop, Groq Whisper transcribes the recording with word-level timestamps. Nemotron then scores clarity, fillers, pacing and confidence, and names one thing to fix.
  • Learned timing. The timestamps are aligned to your script to measure how long you spend on each sentence. Every rehearsal updates a weighted average, so recent runs count more.
  • Present (experimental). Nemotron reads your rehearsal history and writes a three-point briefing. The script then advances on your own timing, with back and forward buttons for drift. Alignment is still being tuned.
  • References. Upload a PDF or text file and set trigger phrases. When you say one, the matching sentence streams into a side panel.

How I built it

Browser (MediaRecorder, Web Speech API) → Netlify serverless functions → Groq Whisper for transcription, Supabase for sessions and sentence timings, and Nebius Token Factory for Nemotron.

  • One function, analyse-transcript, calls nvidia/Llama-3_1-Nemotron-Ultra-253B-v1 through Token Factory's OpenAI-compatible API in three modes: script review, post-rehearsal debrief and day briefing. Each requests a fixed JSON schema at temperature 0.3.
  • API keys live in Netlify environment variables, so the browser never sees them.
  • Token Factory meant no GPUs to provision: one key, one endpoint, one function.

PresentIQ existed before August 26 with rehearsal, Groq transcription, the Supabase timing model and references. During the submission period I added the Nemotron layer on Nebius (Script Doctor, the debrief, the briefing) and auto-scroll.

Challenges

  • Live tracking failed. The voice-following teleprompter was the original product. Abandoning it hurt, but it gave the project its idea.
  • Alignment. Matching Whisper's word timestamps back to script sentences is harder than it looks: repeated words pull sentence boundaries to the wrong place. This is the step I'm still tuning.
  • Structured output. My first Nemotron calls came back empty, and getting consistent, parseable JSON took several rounds of debugging the request.

What I learned

Use the language model for judging and explaining, and use measured timing for the real-time job. Nemotron never has to keep up with a speaker; it reads what already happened. And the unglamorous step, alignment, decides whether the clever part works.

Results so far

Seven rehearsals of one script by one speaker (me): the app measured my pace rising from 61 to 114 words per minute, with 0 to 2 filler words per run. That is one person's practice. It shows the loop is usable.

What's next

  • Use a smaller Nemotron for the fast calls and keep Ultra for the debrief.
  • Highlight the sentences Nemotron flags in the next rehearsal, and set scroll speed from measured pace.
  • Accounts and mobile.

Built With

Share this project:

Updates

Submission history