Privileged & medical files
A law firm handling medical claims can't send patient records to a server abroad. One leak is one too many.
Local 1st AI builder for Israeli businesses
Local AI that never leaves your building. Your files stay yours.
I'm a Local 1st AI builder: I design, install and train local AI models on your own servers, for law firms, defense suppliers, clinics, training teams and any business that can't send its data to ChatGPT. From the first conversation to a model your team actually uses.
Hebrew-first
Runs fully offline
You own the model
Hover to play
0bytes of your data leave your network
11years in the defense industry
6years in high-tech
Why cloud AI isn't an option
Everyone wants AI.
Everyone wants AI.
Not everyone can upload their files to it.
Your competitors paste documents into chatbots. You can't, and you shouldn't. Confidentiality, regulation and plain common sense say your data stays inside. Until now, that meant no AI at all.
Classified & air-gapped
Defense suppliers work on networks with no internet at all. Cloud AI simply can't get in. A local model can.
Data you won't hand to OpenAI
Client lists, HR files, IP, internal know-how. Anything you wouldn't email to a stranger shouldn't be training someone else's model.
A local model solves all three. Here's how
The service
From one conversation
From one conversation
to a private model that works.
One person, end to end. I learn how you work, check what you already own, and only then recommend what to build.
01
Learn the needs
Interviews with the people who'll use it. Which documents, which tasks, what "good" looks like, and what must never leave the room.
02
Audit your hardware
Servers, GPUs, workstations, network, private cloud. Often you can start with what you already have.
03
Recommend a plan
Which open model, what hardware to buy (if any), where it runs, and what it costs to operate. Every model and component is checked for origin and license, including commercial and defense use. A clear plan with options, not a sales pitch.
04
Build the model
Install it, connect it to your documents, and shape it around your real workflows: drafting, search, summaries, Word in and Word out.
05
Train your people
This is where my training background matters most. Hands-on sessions so the team trusts the tool, uses it daily, and knows its limits.
06
Fine-tune when it pays off
When your language, style or domain is unique, I fine-tune the model on your own data, on your own machines.
Steps 4 to 6 in motion
Tune the model, then train the people.
Your documents shape a thin adapter layer on top of an open model. Then the team learns to use it in hands-on sessions. Click the documents, the model or a person to see each part.
↻After launch we measure, refine, and add the next use case. Every loop is faster than the last.
What can run on your hardware
Start with what you have.
Start with what you have.
Buy only what's missing.
Pick the setup closest to yours to see what's realistic. The real answer comes after I check your machines.
Tier 1: Your existing computers
A fast start with nothing to buy
Great for a pilot, small teams and L&D tool builders. Smaller, compressed open models run surprisingly well on everyday machines.
- Models
- Small open models (≈7–14B, quantized)
- Good for
- Summaries, Q&A over documents, Hebrew transcription
- Users
- 1–3 people at a time
- Hardware cost
- ₪0 extra hardware
Indicative ranges. The final recommendation depends on your documents, number of users and required speed.
For the technical reader: CTO, CISO and ITWhat's under the hood
- Open-weights models (e.g. Gemma, Llama, Qwen, Mistral, DictaLM), chosen per task and language
- Local inference server on your hardware or in your own cloud tenant
- Private document index (RAG), with answers that cite the source page
- Fine-tuning with LoRA / QLoRA when your style or domain requires it
- Internal web interface on your network, using your existing user permissions
- No telemetry, outbound traffic blocked at the firewall, logs kept on-prem
What you actually own
Not a black box in the cloud. A folder on your server.
The model, your fine-tuning, the document index and the interface all sit on your own disk. No monthly AI subscription, no telemetry, and it keeps working when the internet goes down.
Glossary
/srv/private-ai/- Open model
Open-weights model - The model itself: a file you download and run on your own machine. Nothing is sent to anyone to get an answer.
- Fine-tuning
LoRA adapter - Extra training on your own examples, so the model writes in your style and terms. It's saved as a small add-on file (an adapter) next to the model.
- Search index
RAG / Embeddings - A smart index of your documents that lets the model find the right page and quote it. Built on your server, from your files.
- Internal interface
Web UI - A chat screen in the browser, available only inside your organization's network.
- Logs
Logs - A record of questions and answers. Kept only on your server, and you decide whether to keep it at all.
- Closed network
Outbound: blocked - All outgoing traffic is blocked. Even if someone tried, nothing leaves the network.
Who it's for
Built for businesses where
Built for businesses where
privacy isn't optional.
Law firms
A private "brain" that knows your past filings, drafts new ones in your firm's voice, and learns from every lawyer's edit.
Defense & security
Models that run on isolated networks with no internet, for document search, technical writing and knowledge retrieval, installed from vetted offline packages. After 11 years in the defense industry, I know how work gets done in these environments.
Training & L&D teams
Local AI for instructional designers and trainers: tools that help build presentations, images, videos and practice scenarios faster, on the internal network. Built the Vibe Training way, without exposing internal content to public models.
vibetraining.co.il ↗Any privacy-first business
Clinics, accountants, HR, R&D, media. If it has to stay separate from the internet or from public LLMs, we can build it.
Local vs. cloud
Same intelligence. Different address.
Worth knowing: the biggest cloud models are still stronger at some tasks. Part of the plan is telling you plainly where a local model is enough, and where it isn't.
A local-first AI builder.
That's what I want to be known for. Everything I build runs in your environment, works on your data, and stays yours. These are the rules I build by, with the projects that prove them.
I share the builds on LinkedIn ↗Your data stays yours.
Documents, recordings and conversations stay with you: on your computer, in your folder or in your private cloud. Nothing passes through someone else's server unless you chose it.
See:Your model, your account.
Bring your own model and your own key. No lock-in to one vendor, and you can swap the model without rebuilding everything.
See:Built into what you already use.
Google Workspace, Obsidian, a shared folder or the office server. Not another system to approve, learn and pay for.
See:Nothing happens without a yes.
Agents can do a lot, but sending, paying or deleting always stops and waits for a person to approve.
See:Hebrew first.
Hebrew and right-to-left are the default, not a translation layer. That includes transcription, summaries and writing.
See:I use what I build.
Every rule here was tested on me first. My agents, my notes and my recordings run this way every day.
See:
What I built.
And what I'm building now.
My own tools and client projects, all built on the same rules. Client names are left out on purpose.
In daily use
Sidebots
My personal AI agent team. A private agent runs on my Mac with an open model and takes tasks from WhatsApp. The browser agent stops for my approval before anything is sent or paid.
Runs on: my Mac, Ollama, Obsidian
In daily use
Anaf
A Hebrew, right-to-left outliner that works as an Obsidian plugin on Mac and iPhone. Every note is a plain Markdown file you own, with no server and no locked format.
Runs on: Obsidian, your own files
Private beta
A Hebrew transcription tool
Records and transcribes meetings in Hebrew. The recording, transcript and summary are saved as files in your own folder, not on someone else's server.
Runs on: Mac, your own folder
Being built now
A private brain for a law firm
A system that knows the firm's past filings, drafts new ones and learns from every lawyer's edit. It all runs inside the firm's own Azure environment.
Runs on: the client's Azure
Built at Elementor
Vibe Training at Elementor
Learning and enablement tools I built inside Elementor's own Google Workspace, with no developers and no new outside system. This is what I talk about in the podcast with Or Daniel.
- An internal learning system
- Automations built on Apps Script
- Interactive learning modules
- Enablement tools for the teams
Runs on: the company's Google Workspace
Live
This website
No cookies, no tracking, and the fonts are embedded. Here too, nothing leaves.
Runs on: a static host
for fun
What I build as a maker
Small ESP32 boards and e-ink screens on my desk. Same rule as everything else: it runs on the device and nothing goes to the cloud.
A pocket book reader
My own firmware for a small e-ink reader built on an ESP32. Books load from my own files, with no store and no account.
Xteink X4, ESP32-C3A 7.5 inch e-paper display
An ESP32 e-paper screen I'm turning into a quiet display for the day. It barely uses power and pulls nothing from the cloud.
reTerminal E1001, ESP32-S3Transcription devices
A small round device for the middle of the meeting table. It records, glows while recording, and pauses when you flip it over. The audio goes straight to the Mac in the room.
AMOLED 1.75", ESP32-S3A digital notebook
Handwriting and notes on e-ink that end up as plain files in my own folder.
ESP32, e-ink
Local AI · Israel
Local 1st AI Builder
Hi, I'm Sagi.
I come from 11 years in the defense industry, a stretch as a freelancer, and 6 years in high-tech.
My background is training and instructional design, and I learned to build. In my last role at Elementor, as an AI Tailor, I led internal AI enablement and building tools that teams use every day. A local model is only worth something if people actually use it, so I build the model and I also train the people who work with it.
- Today: on the founding team of a new startup that brings AI into small and medium businesses
- 11 years in the defense industry
- 6 years at Elementor, most recently as an AI Tailor leading internal AI enablement
- Founder of Vibe Training, a method that teaches training teams to build their own AI tools
- Works in Hebrew and English. Based in Israel, on-site at your office when needed
The video loads from YouTube only after you click.Episode summary on LearnTech ↗
Questions
Everything you'd want to ask.
What is a local AI model?
A local model is an AI model (LLM) that runs on your organization's own computers or servers instead of someone else's cloud. Documents and questions never leave the network, there is no per-user subscription, and it keeps working with no internet.
What does deploying a local model look like?
Four steps: map what the team needs, check the hardware you already have, a written plan with model, hardware and budget, then installation, connecting your documents, fine-tuning when it pays off, and training until the team uses it every day. A focused pilot usually takes a few weeks.
Do I need to buy expensive GPUs?
Not necessarily. Smaller models run on a decent server or even a strong workstation. If speed matters, I'll recommend the right hardware, with a price and a reason.
Does it really work in Hebrew?
Yes. Today's open models handle Hebrew well, and I test every setup on your own Hebrew documents before you commit to anything.
Can it run with no internet at all?
Yes. I prepare offline install packages, they go through your approved transfer process, and after installation everything runs on the internal network. Updates come in the same way, when your security team approves.
How do you check where a model comes from and what its license allows?
Before anything is installed you get a list of every component: the language model, the embedding model, image and voice models and the software around them. For each one: who made it, its country of origin and what the license allows, including commercial and military use. Components that don't meet your requirements, for example by origin, are replaced with alternatives that do.
Can it also make presentations, images and videos?
Yes, with local tools for writing and building slides, generating images and producing narrated video, all on the internal network. Text and images are mature today; local video is less mature, and I'll show you on your own content what works well and what doesn't before you commit.
What's the difference between connecting documents and fine-tuning?
Connecting documents (RAG) lets the model read and cite your files. That covers most needs. Fine-tuning changes the model itself so it writes in your style or domain. We start with the first and move to the second only when it's worth it.
Can it run in our private cloud instead of the office?
Yes. For example, in your own Azure environment in the Israel region, closed to the internet. The data stays in an environment you control.
How long does it take?
A focused pilot usually takes a few weeks from discovery to first real use. Larger rollouts are planned in stages.
How does this connect to Vibe Training and L&D teams?
Vibe Training teaches training teams to build their own learning tools with AI. In organizations where training content can't go to the cloud, I connect the method to a local model on the internal network and train the team to work with it.
Let's talk
Bring AI in.
Bring AI in.
Keep your data in.
A 30-minute call. You tell me what you're trying to do; I'll tell you honestly whether a local model fits.