Your knowledge.Your model.Your walls.
Persto is software you install on your own servers. It turns your organization’s files into a private AI model that runs fully offline and plugs into the tools you already use. Persto, your AI smith, asks plain-language questions in three short stages; you answer, he does the machine learning.
For organizations:Request early access
In development. Early access is open to a small number of organizations.
Meet Persto, your smith. He asks the questions in plain language, you answer, and he forges the model. Illustrative example.
How it works, in three stages
Click through an example workspace.
Upload
Persto
Drop in your folders, any format. I’ll read every file, right here on your server.
| Folder | Files | Status |
|---|---|---|
| policies/ | 214 | Read |
| maintenance-manuals/ | 402 | Read |
| contracts/ | 188 | Queued |
| rate-tables/ | 96 | Queued |
| training-decks/ | 311 | Queued |
| support-inbox/ | 73 | Queued |
Stored on srv-04. Nothing leaves this server.
The documents worth learning from are the ones you can’t send out.
- Cloud AI is off the table
- Policies, contracts, manuals and case files often can’t leave your network, by regulation or by choice.
- Building your own takes specialists
- Preparing data, training and evaluating a model is work for a machine learning team that most organizations don’t have.
Persto lets the people who know the documents build the model themselves, on hardware you control.
What makes Persto different
01
It understands your data before it trains on it
Stage 1 is about your documents. Persto reads every file, whatever the format, shows you what it found, and asks about anything it can’t decide on its own. Then it cleans everything up: duplicates and junk removed, conflicts resolved, sensitive details handled, all in one standard format ready for training.
Data X-ray
A report on the whole collection: what it’s about, main topics, document types, terminology and undefined acronyms, duplicates and versions, possible contradictions, sensitive information, blank or unreadable pages, and how the documents are structured.
Questions about your data
Only the ones that matter, each with a recommended answer. Which version is current? What does this acronym mean? Remove, mask or keep the sensitive information?
One clean, standard structure
Persto removes duplicates, junk pages and repeated headers, applies your answers, and converts everything into one consistent format ready for training.
ACME CONFIDENTIAL · Rev 4 · Page 12 of 401
4.2 Carry-over. Employees may carry over up to 5 days2 of unused leave into the next year.
Requests go on form HR-7, signed by your FSM3.
Questions? Call Dana at (555) 014-22904.
- topic
- Leave carry-over
- rule
- Employees may carry over up to 10 days of unused leave.2
- process
- Requests go on form HR-7, signed by your Field Service Manager (FSM).3
- contact
- HR, [phone masked]4
- source
- leave-policy-2023.docx §4.2
- 1
Repeated header removed
It appeared on all 40 pages and would have taught the model nothing.
Persto, automatically
- 2
Outdated rule replaced
The 2021 policy said 5 days; the 2023 one says 10. You picked 2023 in Stage 1.
Your call
- 3
Acronym spelled out
FSM = Field Service Manager, from the glossary in another file. You confirmed it.
Your call
- 4
Personal phone number masked
Flagged as sensitive in the Data X-ray. You chose to mask it.
Your call
Works with all your file types
- DOCX
- XLSX
- CSV
- PPTX
- TXT
- MD
- HTML
- JSON
- EML
Documents, spreadsheets, slides, text, web pages and email exports, mixed together in one upload.
02
It becomes your model
Stage 2 steers the training and Stage 3 is your sign-off. In between, Persto builds and checks everything itself.
Questions about your goals
Who will use it, how it should write, whether to cite sources, what to do when it isn’t sure. Each answer becomes a training setting. No hyperparameters.
Verified training, on a small open model
Persto writes training examples from your documents, checks every one against its source, and fine-tunes a small open model on them, on your own hardware.
You approve it, then it runs offline
Review sample answers and approve or improve them. Persto retrains on your corrections, then packages the model to run fully offline.
Model report
Tested on 250 questions you provided
- Uses your terminology correctly61%→94%
- Cites the right source document48%→89%
- Says “I don’t know” when it should35%→91%
03
Inside your walls
Persto is software you install on infrastructure you control. Training and the finished model run inside your walls, and the model belongs to you. If you’d rather not run it yourself, Persto can train and host it for you instead.
Air-gapped
For networks physically cut off from the internet. Installed and updated offline, from approved media.
On-premises
On servers in your own building, behind your existing firewall and access controls. Designed for modest hardware.
Private cloud
In your own cloud account (commercial or government), walled off from everyone else, in the region you choose.
Two ways to use Persto
Persto Private
Install it. Everything stays inside.
The core product · required for air-gapped and regulated data
- Software you install on your own servers or private cloud
- Training and the finished model both run on your hardware
- Nothing leaves your environment, and the model is a file you own
Persto Hosted
We train it. You call it.
For teams without GPUs or servers to spare
- Upload your files and answer the same three stages
- Persto trains your model on its own infrastructure
- You get a private endpoint to call from your apps
Either way, your model speaks a standard API, so it plugs into what you already use: internal tools, chat apps, help desks and your own code.
Principles we’re building to
- Your data stays yours
- With Persto Private, documents never leave your environment.
- No outside calls
- Persto Private has no telemetry and no phoning home. Built to work unplugged.
- Every step is logged
- An audit trail of ingestion, cleanup, your decisions and every training run.
- You own the weights
- With Persto Private, the model is a file you keep, independent of any subscription.
These are design principles for a product in development. Persto does not hold certifications such as FedRAMP, SOC 2, HIPAA or CMMC.
Uploading files to a chatbot, compared with owning a model
| Capability | Chatbot with uploaded files | Persto |
|---|---|---|
| Works offline and air-gapped | No | Yes |
| Learns your terminology and formats | Partly | Yes |
| Behaves the same way every time | Partly | Yes |
| Cites its sources | Partly | Yes |
| Proves accuracy on your own questions | No | Yes |
| You own the model | No | Yes |
| No machine learning expertise needed | Yes | Yes |
Works offline and air-gapped
- Chatbot with uploaded files
- No
- Persto
- Yes
Learns your terminology and formats
- Chatbot with uploaded files
- Partly
- Persto
- Yes
Behaves the same way every time
- Chatbot with uploaded files
- Partly
- Persto
- Yes
Cites its sources
- Chatbot with uploaded files
- Partly
- Persto
- Yes
Proves accuracy on your own questions
- Chatbot with uploaded files
- No
- Persto
- Yes
You own the model
- Chatbot with uploaded files
- No
- Persto
- Yes
No machine learning expertise needed
- Chatbot with uploaded files
- Yes
- Persto
- Yes
Typical general-purpose chatbot tools, compared with Persto’s design goals. Persto is in development.
Built for
Organizations whose documents can’t leave the building. Here are a few of the uses we’re designing for.
Defense and government contractors
Draft responses from technical manuals and contract requirements on an air-gapped network.
Healthcare
Answer staff questions from clinical policies and procedures, with citations, inside your network.
Finance
Answer compliance questions from your own policies, offline.
Legal
Draft clauses in your firm’s style from your own precedents and playbooks.
Critical infrastructure
Help technicians find the right procedure in maintenance manuals and safety standards.
Something else
Tell us about it in the early-access form.
Who’s building Persto
Instructions get ignored.Training sticks.
Aaron Pinto
Founder & CEO
- AI/ML engineer at a major defense contractor. Builds agentic AI systems used by 47,000+ employees.
- Amazon (Alexa), at HQ2 from age 18. Built an AI agent on AWS Bedrock and a retrieval pipeline over 8,000+ pages of internal docs; cut bug triage time 70%.
- Fine-tunes models hands-on. Trained a Qwen model with LoRA for an AI portfolio recruiters can talk to.
- Hackathon winner. ROTTED: a Rust vector-search engine for 50 GB document uploads.
- B.S. Computer Science, Western Governors University, 4.0 GPA (December 2026).
I kept watching teams try to build AI agents for their work, and they kept failing. An agent follows your rules only as long as its instructions hold, and under pressure they don’t.
Ask ChatGPT for something dangerous and it refuses. That refusal isn’t a line in a prompt; it was trained into the model. That’s the difference Persto is built on: your knowledge and your rules, trained into a model you own, instead of written into instructions and hoped for.
Request early access
Persto is in development. We’re opening it to a small group of organizations first, and we’ll only contact you about Persto.
Design partners
Have a real use case and a team willing to give feedback? Design partners try early builds in their own environment and help decide what we build next.
Questions
Is Persto available yet?
Not yet. Persto is in development, and we’re opening it to a small group of early-access organizations first. Join the list and we’ll be in touch as places open.
Does my data leave my environment?
Not with Persto Private. It’s software you install on your own infrastructure (air-gapped, on-premises or in your private cloud), and training and the model both run there. With Persto Hosted, you choose to send your files to Persto for training and call the model through a private endpoint. Regulated and air-gapped teams should use Private.
Does it work with the systems we already have?
Yes. Your model is served through a standard API, so your internal tools, chat apps, help desk and your own code can call it, whether it runs on your servers (Private) or on Persto’s endpoint (Hosted).
Do I need machine learning experience?
No. Persto asks plain-language questions about your data and your goals and turns the answers into training decisions. Knowing your documents is the expertise that matters.
What kinds of data does it work with?
All the common ones: PDFs, Word documents, Excel spreadsheets and CSVs, PowerPoint decks, plain text and Markdown, web pages and wiki exports, JSON, and email exports. Mix them freely; Persto converts everything into one standard structure before training. If your data lives in a format we haven’t listed, tell us on the form.
Do I need my own GPUs?
Only for Persto Private, where training runs on your hardware. Persto is being designed for ordinary servers rather than large GPU clusters, and we’ll share specific requirements with early-access participants. With Persto Hosted, you don’t need any: we train and serve the model for you.
What do I get at the end?
With Persto Private, a single model file that you own (the trained model, your settings and an audit log of how it was built), running on your infrastructure behind a standard API. With Persto Hosted, a private endpoint your apps can call. Either way, you get the same three stages and the same testing on your own questions.
What does a design partner do?
Design partners work with us during development: sharing requirements, trying early builds in their own environment, and shaping the roadmap. Tick the box on the form if you’re interested.