Private AI chat
A familiar chat assistant for drafting, summarizing and rewriting. Sign in with Microsoft Entra ID, Google Workspace or on-site Active Directory.
PRODUCTS
One platform. Four sizes. Owned by you.
Every Saddle Tree system runs the same private AI platform. Only the hardware and the models loaded change, so you can start small and move up a tier without starting over.
What your staff get
Staff open the assistant in their browser and sign in with the account they already use. Behind it sits a platform that knows your documents, follows your roles, and keeps a record.
Answers from your own documents
Ask questions of your policy manual, procedures, regulations and files. Each answer cites the source document. Access follows roles: billing staff can't pull HR files.
Automations with human review
Multi-step jobs that run on the box: paperwork and document intake, record and meeting summaries, dictation to a draft, email sorting and draft replies. A person reviews every output before it's final.
Speech-to-text
Turn dictation and recorded meetings into text on your own server, then into a draft in your template.
Document reading (OCR)
Read faxes, scans and PDFs, pull out the fields that matter, and show a confidence score for staff to check.
Admin dashboard
For your IT or security lead: users and roles, connection health, audit-log search, model quality checks, pending updates and approvals, and a compliance checklist report.
Workflows built like software, not screen-scraping
We write automations as tested, versioned code on our workflow engine. Each one has defined steps, uses only the data its job needs, and pauses for human review before anything is final. Your custom workflows are kept separate from the core platform and tested against every platform update, so an update doesn't break the work we built for you.
- Email is drafts only by default. Sending needs a setting turned on for that workflow, plus a person's approval.
- No direct writing into your core systems. The AI prepares the draft, and your staff review it and enter it.
- We don't screen-scrape your systems without written approval from you and the vendor.
- Document arrives
- Read (OCR)
- Extract fields
- Human review
- Intake checklist
Strong open models, tested on your own work
We run leading open AI models on your hardware, chosen and tested for each deployment. No per-token fees, and no data sent to an AI vendor. When a better model ships, we test it on your own examples before switching, and the previous model stays available for rollback.
Two editions, one codebase
Organizations that handle health information get the HIPAA edition. Its safeguards are locked settings that no customization can turn off.
| HIPAA edition | Standard edition | |
|---|---|---|
| Built for | Organizations that handle protected health information | Every other organization that handles confidential information |
| Contract | MSA and statement of work, plus a BAA | MSA and NDA, with the security terms your obligations require |
| Multi-factor sign-in | Required, checked at every login | Strongly recommended |
| Audit log | Every prompt, response, document retrieved, workflow action, admin change and support session. Tamper-evident and append-only | Sign-in, admin and configuration events. Prompt logging optional |
| Cloud AI | Never. Local models only | Off unless you opt in under your own cloud contract |
| Encryption at rest | Full disk, unlocked by the TPM chip plus a network key or PIN, so a stolen server won't unlock off site | Full disk, unlocked by the TPM chip |
| Outputs | Labelled "AI draft: requires professional review." Clinical-decision uses disabled | Labelled as AI drafts |
Remote support is off by default; we connect only when you approve each session. Every session uses multi-factor sign-in and is recorded in your audit log.
Hardware sized to your team
Capacity figures are planning estimates, not measured benchmarks. We confirm sizing with a load test on your real usage during the pilot.
S1 Pilot
Pilots and small offices
- What it is
- A small, quiet desktop AI computer (NVIDIA DGX Spark class, 128 GB unified memory).
- Typically fits
- Up to about 25 named users, with around 5–15 using it at once. Entry system for Standard-edition teams, departments and smaller offices.
- Runs
- The full platform, a mid-size general model, speech-to-text and document search.
- Good to know
- Slower than the closet tiers on large models, with a single drive and no drive mirroring.
M1
Most common
Our standard production box
- What it is
- A tower or rack server for a locked closet. One NVIDIA RTX PRO 6000 Blackwell GPU (96 GB), error-correcting memory, two mirrored drives, a TPM security chip, and battery backup with automatic safe shutdown. A second GPU slot is kept free.
- Typically fits
- Organizations of about 50–150 employees.
- Runs
- The full platform with the default general model, speech-to-text, OCR for faxes and scans, and document search, with room to grow.
- Site needs
- A standard 120 V / 15 A circuit and a ventilated, locked closet.
M2
Larger teams and GPU redundancy
- What it is
- Two of the M1 GPUs (192 GB of GPU memory), a workstation-class processor, more memory and a larger battery backup.
- Typically fits
- About 100–300 employees, heavier multi-department automation, or organizations that want a second GPU as a backup.
- Runs
- Either a larger model, or two copies of the standard model so work continues if one GPU fails. We recommend the redundant setup on the HIPAA edition.
- Site needs
- A dedicated 20 A circuit, plus cooling. A closed closet usually needs an exhaust fan or AC.
L1
By consultation
Larger and multi-site organizations
- What it is
- A four-GPU rack or large tower system (up to 384 GB of GPU memory).
- Typically fits
- About 300–1,000 employees, or multi-site organizations.
- Runs
- The largest open models we support on site, for multi-department or multi-site use.
- Site needs
- A 200–240 V circuit and a real server room.
Teams of about 25–50 people usually start on M1. It has room for heavier scanning, OCR and dictation volume. An office where no more than about 25 people will use the system, with light scanning and dictation, may fit S1. We confirm the size in your written plan.
| S1 Pilot | M1 | M2 | L1 | |
|---|---|---|---|---|
| Best for | Demos, pilots, small offices | Most mid-size organizations | Larger orgs, heavier automation, GPU redundancy | Multi-site or larger orgs |
| Typical size (estimate) | Up to ~25 users | ~50–150 employees | ~100–300 employees | ~300–1,000 employees |
| GPU memory | 128 GB unified (shared with the processor) | 96 GB | 192 GB | Up to 384 GB |
| Drive mirroring | No | Yes | Yes | Yes |
| GPU redundancy option | No | Add a second GPU (becomes M2) | Yes | Yes |
| Placement | Desk or shelf | Ventilated locked closet | Closet with cooling | Server room |
| Power | Standard outlet | 120 V / 15 A circuit | Dedicated 20 A circuit | 200–240 V circuit |
| Availability | Available | Available | Available | By consultation |
| Pricing | Request a quote | Request a quote | Request a quote | Request a quote |
Every deployment is scoped and quoted for your organization. Hardware is sized to your workload and owned by you.
What comes with every box
- Burn-in testing in our lab before it ships
- Asset tag, with serial numbers recorded
- Battery backup (UPS) with automatic safe shutdown
- A sealed spare drive kept on site (mirrored tiers)
- A "who to call" card on the unit
- Full-disk encryption, with the recovery key sealed and held by you
- The manufacturer's warranty passed through to you. We coordinate any repair with the manufacturer.
Connects to the systems you already use
| System | What we connect | Good to know |
|---|---|---|
| Line-of-business systems | Read access through the vendor's API where they allow it, or scheduled report exports. We confirm what your system allows before we build | AI prepares drafts, and your staff review and enter them. No direct write-back. No vendor partnership or certification is implied |
| Fax and scanned documents | A watched folder, a scan-to-email or fax-to-email mailbox, or your e-fax service | Every extraction gets human review |
| Email and calendar | Microsoft 365. Other mail systems through standard IMAP | Drafts by default |
| Documents | SharePoint, OneDrive, network file shares and document management systems, where licensed | Follows your existing folder permissions |
| Accounting and financial | File shares and report exports | Read-only |
| Sign-in | Microsoft Entra ID, Google Workspace, Active Directory | Roles come from your existing groups |
| Notifications | Microsoft Teams or Slack | Messages carry a task link, never sensitive content |
Prefer the cloud? We'll tell you if it fits better.
Sometimes the right answer is a cloud AI service you already use, such as Microsoft's. If that suits you better, we'll say so. We'd rather give you the right answer than sell you a box. Cloud-based engagements are scoped case by case, under your own agreement with the vendor. The on-site safeguards described on this site, such as encrypted drives in your building and the on-server audit log, apply to the on-site system.
Technical details for IT and security reviewers
Technical details for IT and security reviewers
- Models: leading open AI models for chat, speech-to-text, document reading and search, chosen and tested for each deployment.
- Serving: proven open-source model-serving software, matched to each tier.
- Gateway: an internal AI gateway controls which roles can use which models. In the HIPAA edition it permits local models only.
- Guardrails: PII detection for log redaction, plus content-safety checks on inputs and outputs.
- Data: Postgres with pgvector, on encrypted volumes.
- Interface and sign-in: a self-hosted chat interface (LibreChat), with Keycloak brokering your identity provider.
- Monitoring: local Prometheus and Grafana. Only allow-listed health metrics and error codes leave the box.
- Updates: signed images and manifests, pulled over an outbound-only encrypted connection, with automatic rollback.
- Licensing: built mostly on open-source components with permissive licenses.
Not sure which size you need?
That's what the consultation is for. We look at your staff, your workload and your closet, then recommend a tier in writing.