Inspiration
I run several businesses at once, all alongside a full-time job. My best ideas live everywhere. They're in screenshots, notes apps, camera rolls and half-finished drafts and most of them never become anything because I don't have time to sort them.
I wanted one AI I could trust with all of it. Not a generic assistant that checks my calendar, but a private system I could feed my brain to that actually turns my raw life into finished work. Trust was the hard requirement: if it's going to hold everything, it has to run on infrastructure I control.
What it does
SHA-I OS is a private creative operating system for multi-hyphenate creators.
- Inspiration inbox: drop in screenshots and notes and SHA-I tags each one by venture (SHA-I, Namastays, UGC or hackathons) using NVIDIA Nemotron.
- Memory: everything it ingests is stored locally so I can ask things like "what ideas do I have for Namastays reels?" and get answers from my own material.
- Content drafting: SHA-I drafts captions, scripts and post ideas from what's in the inbox.
- Video rough cuts: I pick photos and clips and SHA-I assembles a short reel with MoviePy.
- Approval first: nothing that sends, posts or spends happens without my approval.
This version is intentionally rudimentary. It proves the core loop of capture, understand and create on infrastructure I own.
How we built it
- Models: NVIDIA Nemotron served through Nebius Token Factory using its OpenAI-compatible API. Nemotron Nano handles fast everyday calls like tagging while Nemotron Super handles drafting.
- Model router: every model call goes through one function that reads the provider and model from config, so I can swap models or move SHA-I onto my own hardware later without rewriting anything.
- Privacy: Zero Data Retention is enabled on Nebius so prompts and outputs aren't stored. Personal data and API keys never touch the public repo.
- Stack: Python, SQLite for memory, MoviePy for video and Streamlit for the interface.
- Build process: I planned the architecture with Claude and built it with Claude Code on Windows.
Challenges we ran into
- Locked-down data sources: Google Photos no longer lets apps read a whole library, CapCut has no public editing API and phone notes apps have no API at all. I had to design around each one, using the Photos Picker, MoviePy and folder-based ingestion instead.
- Trust versus convenience: giving an AI broad access is easy. Making it safe and private took real design decisions about what leaves my system and what never does.
- Scope: I wanted to build everything. Cutting down to a working core loop under deadline was the hardest part.
- Learning curve: this was my first time building a personal AI system end to end, from API setup to architecture.
Accomplishments that we're proud of
- A working end-to-end loop from raw screenshots and clips to on-brand drafts, running on NVIDIA Nemotron and Nebius
- An approval-first design where SHA-I can't act on my behalf without my say
- A portable architecture where no single company owns SHA-I or my data
- Building something I'll actually use every day, not just a demo
What we learned
- How to work with open models through an OpenAI-compatible API, and why routing every call through one layer keeps you flexible
- How different "open" and "private" really are, and how to read a provider's data policy
- That small fast models handle most everyday tasks, so bigger models can be saved for real reasoning
- That I have a lot more to learn on my path to becoming an ML engineer, and SHA-I is going to be my classroom
What's next for SHA-I OS
- A full trust dashboard with a complete action log alongside the approval queue
- Daily agents for my marketing brief and an opportunity radar for hackathons and brand deals, with live web search through Tavily
- A voice profile so every draft sounds like me
- Fuller video drafts that export to real editors for finishing
- Moving SHA-I onto my own home machine so my data never leaves me
- Eventually, opening it up to other creators who juggle multiple brands
Log in or sign up for Devpost to join the conversation.