I built this shit

Yeah, I built this shit.

Not a SaaS landing page.
Not yet another 'AI wrapper'.
An actual piece of hardware that sits in the palm of your hand and does what everyone wants.

Zaubra? ChatGPT for your files that lives on your desk, not on someone else’s server.

Fully offline. Private. No bullshit.

What the hell is Zaubra?

Imagine if Shazam worked for your documents.

You plug a small box into your laptop via USB-C, drop in your contracts, PDFs, filings, emails, spreadsheets, whatever and then just ask it questions:

'Summarize this 80-page contract and tell me where I’m getting screwed.'
'What changed between version 3 and version 7 of this agreement?'
'Find every clause related to termination in these 40 PDFs.'
'Explain this German legal text to me like I’m five but still liable.'

And instead of forwarding your life to some US-hosted black box, everything happens locally:

Embeddings? Local.

Retrieval? Local.

LLM inference? Local.

Your data? Never leaves the device.

No 'we care about your privacy' banner. It just… doesn’t upload.

Why I bothered building hardware in 2025

Most devs today are doing prompt gymnastics on top of someone else’s API and calling it a product.

Cool. Have fun.

But if you work in law, finance, healthcare, government, just upload all your client files to our cloud lol' is not a serious sentence.

The reality:

Law firms deal with stuff that is absolutely not meant to sit on a random US server.

Banks & finance have compliance people whose full-time job is to say 'no' to everything with an API key.

Enterprises have DPOs who hear 'AI SaaS' and instinctively reach for the DPIA template.

So instead of trying to talk privacy theater, I went the other direction:

'Here. This box. Your office. Your jurisdiction. Your power cable.
Pull the internet plug if you’re paranoid. It still works.'

That’s Zaubra.

What’s inside the box?

This isn’t a Raspberry Pi duct-taped to a fan.

Under the hood:

NVIDIA Jetson Orin NX – 16 GB of RAM, plenty of CUDA cores.

Local LLM – running quantized models tuned for retrieval + instruction following.

Custom RAG pipeline – optimized for legal/financial documents, page-aware and citation-aware.

Microcontroller brain (STM32) – handling low-level control, boot, monitoring.

USB-C connection – shows up with a companion desktop app; plug-and-play.

No server to 'spin up'.
No kube cluster.
No 'region' to pick.

Your 'backend' is literally sitting on your desk.

What it actually does (for non-PMs)

When you drag a bunch of PDFs into Zaubra, this is what happens:

Ingestion & OCR
- Pages are parsed, OCR’d if needed, cleaned up.
- Per-page text is stored so we can always trace answers back to source.
Chunking with traceability
- Documents are sliced into smart chunks with metadata: filename, page, hash, last modified.
- This isn’t 'random chunk soup'; it’s legal-grade traceable.
Vector + keyword indexing
- Hybrid search: embeddings + BM25-style keyword fallback.
- This means it doesn’t fall apart when the doc is weird, dense, or highly technical.
Query & retrieval
- You type something like:

'Find all clauses that mention jurisdiction and arbitration in these 12 contracts.'

Device pulls the most relevant chunks, with the actual page references.

LLM response
- The model, running locally, generates an answer with citations:
  - 'This comes from Contract_A.pdf, page 12.'
  - 'This conflict is between Document_B page 3 and Document_C page 9.'
All of this happens without touching the internet.
- You can literally firewall the machine. It doesn’t care.

That’s it. No dark magic. Just actual engineering instead of 'wrap OpenAI and ship.'

Who this is for (and who it’s not)

Zaubra is for you if:

You handle sensitive documents you can’t just upload to random clouds.
You need traceable answers with page-level citations.
You like the idea of ChatGPT-like power but under your own roof.

Zaubra is not for you if:

You’re happy pasting client contracts into free web UIs.
You think 'privacy' is just a cookie banner.
You want a to-do list app with an AI sticker on it.