Download for macOS
Built Locally #1: this week’s top local AI apps built on Neywa
This week’s highlighted Local AI apps:
- Local scene analysis for AI video creators, by Sinan Binay
- Hands-free consultation notes, kept inside the clinic, by Clément Pillette
- AI storyteller for kids that keeps conversations at home, by Akash Deb
What exactly can you build with Local AI? Our weekly Community Newsletter answers this directly by featuring Local AI apps built and deployed by developers, builders and entrepreneurs.
AI video production: turn past videos into a reusable library of characters and scenes
Problem
If you make videos of any kind using AI image and video generators, you know the pain of not being able to reuse the creative data from past videos in new projects. Characters, wardrobe, locations, camera work, scene structure, editing choices, dialogue and sound get lost and are never repurposed. You have to describe everything from scratch for every new project.
Solution
Video creator Sinan Binay solved this with Sinners Film DNA, a local-first analysis workspace that extracts this creative data from any type of video: Reels and other short-form social content, ads, product videos, music videos, short films and feature-length films. It converts what it finds into editable source material for future projects. You can use it, for example, to bring back a character from an earlier film, then age it or change how it behaves.
Why Local?
Running locally is a major unlock to this new capability. A single project runs from 14 minutes to two hours, and analyzing it generates a large volume of frames, text and audio. Sending all of that to cloud vision models, again and again, is prohibitively expensive for most creators. Locally, a creator’s creative history also stays private: the original media, extracted frames, analysis results, search index and project files remain on the Mac by default.
How to build
Sinan built it on top of Neywa Labs’ inference engine. The system first detects scene boundaries and selects representative frames from each scene. A vision model then reads a scene’s frames together, in timeline order, so it can follow the same character or object across shots and understand how the space is used. Each scene comes out as structured text covering all of the video’s creative elements. Everything goes into a local search index, from which the tool generates production prompts for new projects and hands material off for editing.
Healthcare: write consultations, pull records and draft prescriptions without touching the keyboard
Problem
If you’re a vet, a dentist or a doctor, much of what you do goes to the screen: typing notes, looking up the last blood test, opening the calendar for a follow-up, writing the prescription or referral letter. Every minute on the keyboard is a minute not spent with the patient in front of you.
Solution
Clément Pillette, a Netherlands-based founder with 22 years in the semiconductor and chemical industries, is building an ambient assistant that takes keyboard and mouse out of the consultation. The vet just talks with the pet owner. The assistant follows the conversation and acts as it goes: it brings up the last blood test when it’s mentioned, opens the calendar when a follow-up comes up, and drafts the prescription or referral letter. When the consultation ends, it writes the summary, and the vet approves every critical action on a review page before anything is final.
Why Local?
Everything runs locally from day one. Nothing is ever recorded: speech is transcribed on the fly, and nothing leaves the clinic. That matters when the conversation is about someone’s health, and it makes GDPR compliance much simpler. Local is fast, too: on one of his test workflows, a frontier cloud model took about 12 minutes and his local setup with GLM-5.3-Flash took three.
How to build
Under the hood, audio is transcribed live, in whatever language is spoken, by open speech models such as Parakeet and Nemotron 3.5 Transcribe. Speaker separation tells the system who is talking. When it isn’t confident, the agent works out who said what from the meaning of the sentence.
Kids’ audio players: turn story cards into characters your child can talk to, with every conversation kept at home
Problem
If your kids love screen-free audio players like Yoto or Toniebox, you are aware of one clear limit: every story is fixed on a card, and the player can’t answer a question or adapt to your child. AI voice toys could fix that, but many parents don’t want their child’s voice and conversations sent to a company’s servers.
Solution
OpenToys, by Akash Deb, is an AI voice companion for kids that turns those players into something personal. Kids talk to the device and ask it for a story, or chat with characters such as Professor Octo, and parents can create new characters or clone a voice. Parents can also make cards on any topic: a family visiting a new city can create a card about the Tower of London, and the toy tells the kids about its history, people and animals.
Why Local?
Child safety is the clearest non-negotiable in AI adoption. Running locally is what answers the parents’ main concern. In the local setup, no data goes to the cloud: conversations, stories and voices stay on the family’s Mac. For European customers, local lets you pick speech models that are strong in French, Italian, Serbian and German as well as English.
How to build
Akash started the local version at the end of last year, building on MLX after finding Neywa Labs’ multimodal engine. The toy runs on an ESP32 chip and connects to a Rust app built with Tauri, which runs the models on the Mac with MLX-Audio. A language model, such as Qwen2.5 at 4-bit, generates the conversation, and a text-to-speech model gives the character its voice. He calls Qwen3-TTS “a game changer” and uses Voxtral for European languages.
Build yours with Nativ
Each of the above products are introducing brand new capabilities, creating genuine value, and started on a Mac with Neywa’s engine.
We built Nativ as the beginning of your Local AI journey: you can iterate invent new products and no longer be chained by limited compute, tight budgets, capped runtime.
