Dikshant Rajput

Just ask, "What's happening in sample.mp4?" or "Can you describe this video?" and it analyzes the content, identifies people, objects, actions, and scenes, then gives you a natural language summary. Simply mention the filename and it finds the video automatically. If it can't locate it, it asks you to pick the file. Once loaded, the video stays in context so you can keep asking follow up questions without reselecting it.
Watch it here: https://youtu.be/Jl4lOjTOGp4

LocalClicky can now edit videos hands free. Just say what you want - "trim my vacation clip from 10 to 30 seconds", "mute the audio in intro", "speed it up to 2x" - and it's done. Name the file and it finds it automatically. If it can't locate it, it asks you to pick it. The file stays in context so you can chain edits back to back without repeating yourself.
Watch it here: https://youtu.be/HM6sD69cQ5Q

The last piece that needed internet was the wake word listener. It was using Google Speech Recognition. Worked fine, but broke the "nothing leaves your machine" promise.

Replaced it with openWakeWord - an open source, on device wake word library. Runs a tiny ONNX model locally, processes 80ms audio chunks through sounddevice, fires a callback when confidence crosses a threshold. Zero network calls.

The tradeoff: no pretrained "computer" model exists yet, so the default is "hey Jarvis". You can train a custom keyword using their pipeline if you want something specific.

Repo: https://github.com/dikshantrajpu...