Lars Jensen

I wanted to see how far a fully local speech-to-text pipeline could go on a laptop — no cloud, no API keys. Whisper Notes records a voice memo, transcribes it with a quantised Whisper model, then asks a small local LLM to clean it up into structured markdown.

What it does

• Records audio and transcribes it offline
• Cleans filler words and punctuates automatically
• Produces a short summary and a list of action items

What I learned

The transcription was the easy part — the interesting work was prompting a small model to reformat messy speech without inventing content.

The smallest useful RAG I could build in a weekend. Point it at a folder, it embeds every document, and you can ask questions that are answered with citations back to the source passages.

What it does

• Watches a folder and embeds new files automatically
• Retrieves the most relevant chunks for each question
• Always shows its sources so you can verify the answer

The goal was honesty: never answer without showing where it came from.

A playful experiment in multimodal prompting. Pixel Poet looks at an image and writes a few lines of verse inspired by it — you can nudge the tone from austere to whimsical with a couple of dials.

What it does

• Describes the mood and details of any image
• Lets you steer style, length and tone
• Keeps a small gallery of poems you liked