Rules and permissions
Before SARAH does anything, the harness asks whether it is allowed, for this person and this surface. When a response would cross a line, it goes to a person.
Home / Capabilities / Architecture
ArchitectureThe model is the engine. The harness is the car.
Everything on the front page runs on this. We do not manufacture the chips, and we do not manufacture the base model. We operate the GPUs, and we host the language model, the speech recognition and the voice. Around them we built the harness: the memory, the rules, the connectors and the private network that make the intelligence something a business can put its name behind.
The stack
One harness. Many governed surfaces.
No army of agents anywhere in the picture: channels arrive at one voice appliance, one memory harness and one connector hub.
ChannelsHow people arrive
Browser link (WebRTC)SIP handsetPublic number as a SIP URIWeb chatWhatsApp and TelegramEmailVideo conference
↓
The voice applianceOur own code, on GPUs we operate
Carrier-grade switchPBXWebRTC gatewaySIP gatewayConference bridgeStreaming speech recognitionLanguage modelSpeech synthesis
↓
The harnessWhat makes it trustworthy
One memory, five tiersRole management per extensionRules and permissionsVerification and fallbackAudit ledger
↓ Layer-2 Private Enterprise IP Network (PEIPN)
SOPHIALayer-3 API hub
CRMCalendarERPPaymentsTicketingProperty managementGoverned gateway to the outside, only where necessary
The voice path
Three engines, all hosted by us.
Language, speech recognition and speech synthesis. None of the three is a public API, so the call does not stop when a public model returns a rate limit or changes its terms.
The switch, the PBX, the WebRTC gateway, the SIP gateway and the video conference bridge were written by us, from scratch. They are not built on FreeSWITCH, Asterisk or any other open-source system. SARAH lives inside the switch: the conversation, speech recognition and speech synthesis run in the same box, so nothing travels and nothing waits.
- Speech recognition streams during the call; it is never batch on a live call
- The audio stays with the customer's record, not with a model vendor
- Calls transfer between AI and human extensions over SIP, in both directions
- On premises, the appliance can run with no internet connection at all
- A 235-billion-parameter model holds the conversation; a 744-billion-parameter model carries the orchestration
The memory
A memory with a soul: five kinds of remembering.
A model is intelligence without a past. SARAH is built the other way around: one continuous, sovereign recollection of every conversation, kept on hardware that belongs to its owner.
Memory is a circulation, not storage. Six loops run under every conversation: she takes everything in, gives each memory a meaning, reflects experience into understanding, recalls by meaning scoped to the person, reinforces what matters and lets the rest soften, and keeps educating herself. The system improves without a single weight being retrained.
- Raw. Every interaction, verbatim, in the language it was spoken.
- Episodic. What happened, in order, to this person.
- Semantic. What is durably true, de-duplicated, with contradictions resolved to the latest truth.
- Procedural. How to do things well: what worked, kept as skills.
- Identity. One person across every channel, tied to their organization and history.
Harness engineering
I
II
III
IV
V
Roles
The five walls of the harness.
Prompt engineering tunes the question. Context engineering feeds the model. Harness engineering builds the world the model lives in.
Tools and skills
An answer is not work. Through SOPHIA, SARAH reads the state of your systems and acts in them, chaining tools to finish a task end to end.
Verification and fallback
Meaningful output is checked against a golden source: your records, knowledge base and policies. If it cannot be confirmed, it falls back to a stricter path or to a person.
Context and memory
Session context for the conversation in front of her and a persistent memory for the relationship, so every interaction begins where the last one ended.
Audit and feedback
Every action is logged: who, what, when, to whom and why. Omega Leads watches which outreach converts and feeds the lesson back.
Many personalities, governed
A personality is a governed configuration: the prompt, greeting, voice, knowledge scope, connector scope and escalation for each extension, editable in the portal without a developer.
The private network
Not exposed to the public internet.
Each client reaches the intelligence over a Private Enterprise IP Network, PEIPN, with its own dedicated VLAN, isolated from every other site on the same Layer-2 fabric.
Voice, memory recall, connector calls and audit events travel governed private paths. The separation is built into the network itself, before any software or firewall. This is what we mean by built to be compliant: the architecture is designed so that exposure does not happen, rather than a badge on a pricing page.
Where it runs
Hosted customers run on our Chicago compute. SARAH Mini, SARAH Pro and SARAH Enterprise can sit on the customer's premises. The Last Mile Mini AI data center is a separate product: our own building, OMEGA Hyperscale.
Models run at full 16-bit precision, not compressed to 8-bit or 4-bit to save cost.
SOPHIA
Heavy lifting, done through your policy.
SOPHIA is the integration and identity layer. SARAH does not need an agent per bank, airline or government agency. She needs one hub and role-scoped connector calls.
Three generic synthesizers build new connectors on first use: REST from an OpenAPI specification, SOAP from a WSDL, and screen automation from XPath selectors for systems with no API at all. Every number keeps the noun it earned.
34,792,085Addressable endpoints and features
1,512,660+Connectors, pre-generated permutations
191,773Categories
1,800First-class connectors, hand-written and tested
Ask where the voice runs.
If the language model, the recognition and the voice are a public API, the rest of the suite is a rental.