Products / Platform
S.A.L.T CI Platform
Sovereign enterprise AI that runs on your own servers, air-gapped by default, under your own brand. One core, a ring of shared AI servers, and twelve applications built on both. Built in Malaysia.
Sovereign AI, without stitching it together
Government agencies, universities, defence and regulated enterprises need AI they control. Until now that meant either sending data to a foreign cloud, or assembling borrowed open-source parts and hoping they keep working together.
Data leaves the country
Popular AI services run in foreign clouds under foreign law. Sensitive records cannot follow them there.
A stack of borrowed parts
A model runner, a chat screen, a vector store and pipelines: separate products to install, upgrade, secure and brand.
Someone else's brand
Licence terms limit rebranding at scale, so the organisation and its partners cannot put their own name on the service.
Four words, one promise
Scalable
From a single notebook to a multi-site national deployment, on the same platform.
Advanced AI
Local models first, any major provider as an optional connector. Generative and agentic.
Local
Data never leaves your infrastructure. Fully on-premise and air-gap capable, in Bahasa Melayu and English.
Trustworthy
Full audit trail, role-based access, no lock-in. Open-source foundations credited openly.
Like salt in food
S.A.L.T CI enhances people without replacing them. It becomes the quiet essential in daily operations, where AI advises and people decide.
S.A.L.T CI in one line
One sovereign AI platform that runs on your own servers, works with the network cut, carries your brand, and grows from one notebook to a national deployment without changing product.
Generative plus agentic, bounded by your rules
S.A.L.T CI brings content creation and autonomous action together, but always inside limits the organisation sets and can inspect.
Creates content
- Text, images and code
- Answers and follows instructions
- Waits for a prompt
Takes action
- Executes tasks and decides
- Works toward goals
- Acts without being asked
S.A.L.T CI
- Learns your context and workflow
- Acts inside a governed blueprint
- Secure, auditable and local
Three rules every S.A.L.T CI product follows
AI works inside the process
Processes are written as blueprints. Agents draft, check, summarise and nudge, but cannot step outside the blueprint.
Never a silent fallback
An air-gapped site never sends a prompt to an external provider. Cloud models are used only where the organisation allows it, and are labelled.
Every action recorded
Requests, sign-ins, changes and tool calls go into an audit log chained by hash, so any alteration or deletion shows.
AI with a dial, not a switch
People do the step.
AI proposes; a person writes.
AI drafts; a person edits and approves.
AI completes; a person reviews after.
AI completes within the rules.
Each AI step has its own level. Low-confidence or uncited output always goes to a person.
Built for people who must answer for the result
Officers, lecturers and engineers stay accountable for what leaves their desk. S.A.L.T CI shows its sources and which model answered, so a person can check the work before signing it.
Three tiers. One platform.
Most AI products are one application. S.A.L.T CI is a core, a ring of shared AI servers, and the applications built on both, so every new capability is built once and reused everywhere.
TERAS is the brain
Runs the models, holds the knowledge, enforces identity and policy, and records every request.
Servers are the senses
Hearing, sight, voice, documents and security, built once as MCP servers and called by every application.
Apps are thin and fast
Each application is a focused front end on top of the platform, quick to build and simple to replace.
How every request travels
Console, MOM Studio, TERUSAN, any OpenAI-compatible tool
Identity, tenant rules, classification and quotas
Local model, linked site, or a provider only if allowed
Audit, usage and cost, chained by hash
One core that runs models, knowledge and policy
TERAS is S.A.L.T CI's own core engine, written from scratch. It installs once and serves every application, and the same platform runs a notebook demo or a ministry-wide rollout.
Multi-model
Local models first; switch models mid-session; 10+ providers optional.
Any hardware
NVIDIA, AMD, Intel, Apple silicon or CPU, mixed in one site.
Knowledge with citations
Upload documents and query them; answers show their sources.
OCR in BM and English
Scanned papers and images become searchable knowledge.
Voice and images
Speech to text, read-aloud and image generation, all local.
Personalisation
Memory and writing style that follow each person across models.
Model Studio
Test sets, judge, scorecards and LoRA fine-tuning before people rely on a model.
Agents and tools
MCP and OpenAPI tools, schedules, webhooks and channels such as WhatsApp and Teams.
Sign-in and access
SSO with Entra ID, Google, Keycloak, LDAP or SAML; roles, keys and quotas.
Hash-chained audit
Who, what, which model, where it ran, the result and the tokens.
Linked sites
Several sites share models and compute, with failover between nodes.
Two-level white label
Client branding and reseller branding, with the framework credited.
Compatible with what is already built
| Interface | For |
|---|---|
/v1/chat/completions · /v1/responses | OpenAI-compatible apps and agents, with streaming and tools |
/v1/messages | Apps written for Anthropic's Messages API |
/api/chat · /api/tags | Ollama-compatible tools, unchanged |
/v1/embeddings · /v1/rerank | Search vectors and re-ranking from local models |
/mcp · webhooks | TERAS as an MCP server for other agents; signed events for other systems |
Full detail in the S.A.L.T CI TERAS product brief.
Prove a model, then teach it your way
Most platforms let you pick a model and hope. TERAS lets the organisation compare, grade and fine-tune models on its own questions, on its own GPU, before people rely on them.
Compare side by side
Ask once and compare up to three models, each with its full trace: route, first-word time, speed and tokens.
Grade on a test set
Build test sets from real questions. A judge model scores every answer, and scorecards show which model fits which job.
Teach it the house style
Train a LoRA adapter on your own dataset and GPU. TERAS tests it before and after, then serves it like any other model.
From dataset to a tested model, in one panel
Import, take from chats, or generate from a knowledge collection
A LoRA adapter on a workstation or server GPU
To the runtime format and added to TERAS
Before and after on the same test set
Many adapters share one engine
Any dataset format
OpenAI messages, ShareGPT, Alpaca, chats, test sets, or generated from documents.
Small GPUs welcome
Small models train on a workstation card; larger ones on servers.
Red-team test sets
Moderation and red-team sets check a model before release, with PERISAI.
A real example: 0% to 100% on its test set
A small model learned an organisation's helpdesk answer format from 61 examples in a training run of minutes on a workstation GPU, then passed its full test set where the base model passed none.
From one notebook to a national network, without a single script
Multi-server AI is usually a project of its own, with load balancers, sync scripts and command-line setup. In TERAS, nodes, compute agents and linked sites are added from the admin panel, without reinstalling.
One install is one site, fully air-gapped if required.
Each node has a full copy; failover by itself with three or more.
Spare machines lend engines, any GPU or CPU, keeping only model files.
Independent sites link with a one-time code, each choosing what it shares.
What a link between sites can share
| Share | What it does |
|---|---|
| Models | All or chosen models, shown as site/model; cloud models only while the giving site allows them |
| Engines (burst) | The other site's engines answer when a model is missing here, when this site is busy, or both |
| Knowledge | Collections searched where they live; documents are never copied |
| Backup space | Holds the other site's encrypted backups, which it cannot read |
| People | Recognise each other's people, so chats and memories follow a person |
Link with a code
A one-time code plus a matching check code on both screens; requests between sites are signed.
Encrypted backups
Scheduled AES-256-GCM backups to a folder, another site, S3, Google Drive or OneDrive.
No downtime
Drain or remove a node, or hand over the main role, while people keep working.
Example: a university that starts small
Lecturers first run TERAS on their own notebooks. Faculty servers join as the main site for people without one, and as burst capacity for those who have one. Later a campus data centre becomes the main site, with every earlier install still linked.
Shared servers that every app can see, hear and use
Shared capabilities are built once as MCP servers and reused by every application, instead of each app rebuilding its own transcription, vision or document engine. Fourteen run today, and the set is open-ended, so a new server is added whenever a client needs a new capability.
Hear, see and speak
- Audio MCP: the ear, speech transcription
- IRIS Vision: the eye, documents, photos and cards
- LOGHAT: the tongue, eight Malaysian dialects
- Avatar Companion: 3D avatar with lip-synced voice
Produce real outputs
- Arch MCP v1 and v2: brief to floor plan, 3D and PDF
- DocConvert: HTML to PDF, DOCX, PPTX and XLSX
- Workspace MCP: generates ready workspaces for teams
Defend and comply
- TINJAU: reconnaissance, scanning and OSINT
- TAHAN: active defence and auto-ban
- JEJAK: forensics and incident response
- PATUH: ISO 27001, NIST, RMiT, PDPA
- PERISAI Hub: unified SOC command
Watch and act
- Surveillance MCP: the SENTINEL detection engine
- Kali worker: runs the security toolchain in the background with no public port
Build once, reuse everywhere
MOM Studio, TETAMU and TERUSAN call the same ear, eye and document servers. Improve the eye once, and every application that reads a photo gets better the same day. Each runs as a container alongside the apps it serves.
Example: one meeting, three servers and the core
Transcribes the recording
Reads whiteboard photos and business cards
Drafts the minutes in the department format
An editable Word draft for the secretary
A new enterprise AI app in days, not months
Every application starts with the platform underneath it, so the work goes into the problem, not the plumbing. This is how S.A.L.T CI ships custom applications for clients quickly.
To scaffold a new MCP server
To wire in a knowledge base
To a branded PWA front end
Additional AI licence cost per app
What every application inherits
Two sandboxes beside the platform, never shipped to clients
REKA Sandbox
An AI coding agent that writes and runs real code in isolated containers on S.A.L.T CI infrastructure, with project memory and access to other S.A.L.T CI products as tools.
S.A.L.T CI Sandbox
A private test engine that speaks the same APIs as real payment gateways, mail servers and WhatsApp: a dummy bank, seven gateway formats, a mail catcher and a WhatsApp look-alike.
From idea to the client's own core
In REKA
Against the Sandbox
Change base URL and keys
On the client's own core
Twelve applications on one platform
Purpose-built front ends that share one gateway, one set of AI servers and one security model. Clients take the applications they need and add more later without changing the core.
| Application | What it does | Status |
|---|---|---|
| MOM Studio | Meeting intelligence: recordings and photos to minutes in each department's own format | Live |
| TERUSAN | Governed process flow: forms and approvals run as blueprints with AI working inside them | Showcase |
| TETAMU | Hospitality platform: hotel operations and an AI concierge on the same records | Showcase |
| PERISAI | Cybersecurity suite: reconnaissance, active defence, forensics and compliance | Live |
| SENTINEL | Intelligence surveillance: fire, faces, plates and loitering across existing cameras | Live |
| S.A.L.T NEXUS | Server and desktop operations: SSH, fleet health and remote desktop with an AI operator | Live |
| IRIS Vision | Vision intelligence: documents, scenes and business cards, local-first | Live |
| LOGHAT | Dialect voice studio: speech in, reasoning, speech out, in eight Malaysian dialects | Live |
| Arch Intelligence | Architectural AI: a written brief to floor plans, 3D massing and a design report | Live |
| NAKHODA | Sovereign browser companion with S.A.L.T CI as the only backend | Live |
| REKA | Agentic build suite for developers on their own infrastructure | In progress |
| PANGLIMA | Sovereign desktop companion that reads the screen and works in desktop apps | In development |
Showcase means a working build demonstrated to clients and ready for a proof of concept. Product briefs are available for MOM Studio, TERUSAN, TETAMU, PERISAI, SENTINEL, S.A.L.T NEXUS, IRIS and TERAS.
Start with one application, grow on the same core
Every application runs on the same TERAS core and shared servers. Adding one later means no new AI stack, no new sign-in and no new audit system: the core is already there.
Where the platform meets daily work
Four applications that turn the platform into outcomes for administrators, process owners, operators and security teams.
MOM Studio
- Recordings, photos of whiteboards and business cards in; minutes out
- Each department's own Word layout: header, logo, tables and numbering kept
- Fair job queue, a dedicated minutes model and editable DOCX drafts
- Each user sees only the templates they are allowed to use
TERUSAN
- Any process written once as a versioned blueprint
- 13 building blocks: tasks, reviews, sign-off panels, rules, scoring, documents
- AI with a dial from Off to Auto on every step
- Append-only, hash-chained case history for every decision
TETAMU
- Rooms, rates, reservations, guests and staff on one core
- An AI concierge that plans halal-aware, weather-aware trips
- The owner sets the bounds: AI can prepare a booking but never confirm or charge one
- Each property gets its own copy on its own server
PERISAI
- 60+ security tools orchestrated by AI
- Red team, blue team, forensics and compliance in one suite
- ISO 27001, NIST CSF 2.0, RMiT, PDPA and CIS v8 mapping
- Self-run software for the organisation's own team
People stay in charge
Across every application, AI output is a draft or a recommendation. Minutes are edited before release, approvals are signed by people, bookings are confirmed by staff, and security actions follow the organisation's own rules.
Intelligence where people already work
On servers and remote desktops, across cameras, in the browser and on the desktop, all powered by the organisation's own S.A.L.T CI gateway.
S.A.L.T NEXUS
SSH shell, agentless fleet health and remote desktop in one web console. NEXUS AI explains, suggests or runs commands; risky steps wait for approval.
SENTINEL
Real-time computer vision on existing cameras: fire, fauna, faces, number plates and loitering, with duplicate alerts suppressed and no cloud.
IRIS Vision
Reads documents, whiteboards, scenes and business cards on local vision models. An app in its own right and the eye of every other app.
LOGHAT
A dialect voice studio: speech in, S.A.L.T CI reasoning, speech out, in eight Malaysian dialects, with an own-voice option.
Arch Intelligence
A written brief to a complete building design package: floor plans, 3D massing renders and an A3 PDF report.
NAKHODA and PANGLIMA
A browser companion for writing and context on any page, and a desktop companion that works in Word, Excel and Outlook.
Where each one runs
| Application | Runs on | People reach it through |
|---|---|---|
| S.A.L.T NEXUS | The organisation's own server | A web app on any device; an agent for desktops behind NAT |
| SENTINEL | A server on the organisation's own network | Alerts and a monitoring console |
| IRIS Vision | The S.A.L.T CI site | A camera-ready web app, and other apps as a building block |
| LOGHAT | The S.A.L.T CI site | A mobile web app, with continuous conversation mode |
| Arch Intelligence | The S.A.L.T CI site | A web app that returns a PDF design package |
| NAKHODA | The person's browser | A browser extension with S.A.L.T CI as its only backend |
| PANGLIMA | The person's desktop | A Windows desktop companion, in development |
Sovereign by design, accountable by default
Security is built into the core, not added per application. Every application inherits the same identity, policy, audit and routing rules from TERAS.
Data classification routing
Each classification decides where a request may run, which models may see it, and whether it may leave the site.
Air-gap enforced
Cloud use needs hybrid mode and the person's permission; cloud answers are labelled.
Data loss prevention
IC numbers, account numbers and secrets masked, warned or blocked per route.
Injection screening
Documents, web pages and tool results checked before they reach a model.
Strong sign-in
Argon2id, authenticator codes, SSO, LDAP, SAML and SCIM provisioning.
Signed between sites
Requests between sites are signed, time-limited and usable once.
Tamper-evident audit
A hash-chained log of every request and change, with a Verify chain action.
Security that watches itself
PERISAI's defence, forensics and compliance modules can monitor and report on the S.A.L.T CI platform itself, so the organisation sees the same evidence an auditor would.
Licence and branding
A licence that never locks people out
Each site carries a signed licence, activated online or by an offline file for air-gapped sites. An expired licence never locks people out of their own data.
Two-level white label
Client branding with S.A.L.T CI credited, or full reseller branding with a small "Built on S.A.L.T CI Framework" notice.
Runs on whatever you already have
Not tied to any machine, vendor or size. New and older cards, different vendors, GPU and CPU machines can work together in one site.
Personal & edge
Notebooks, workstations and rugged machines, fully offline.
Servers & VMs
Bare-metal x86 or ARM; VMware, Hyper-V, Proxmox and KVM.
Sovereign cloud
Sovereign or private cloud with containers or Kubernetes.
Sizing guide
One machine, models up to about 8B parameters.
One server with one or two GPUs, models up to about 32B.
Multi-node GPU site plus compute agents; 70B and larger.
Several linked multi-node sites sharing compute.
Indicative only. Final sizing is confirmed during proof-of-concept load testing.
Hardware
| Hardware | Engine | Notes |
|---|---|---|
| NVIDIA GPUs | CUDA | x86 and ARM; one or several cards |
| AMD GPUs | ROCm or Vulkan | Data-centre and desktop cards |
| Intel GPUs | Vulkan or SYCL | Discrete and integrated |
| Older GPUs, any vendor | Vulkan | A second life as compute agents |
| Apple silicon | Metal | As a site or a compute agent |
| Any CPU | CPU | Every feature works, more slowly |
Install in minutes, upgrade in one step
A Windows installer with a system check, one script for Linux and macOS, a Helm chart, and an offline bundle for sites that never connect. Upgrades take a backup first.
Proven on client servers, with the network cut
Every proof of concept to date ran on-premise with local models, including tests with the internet fully disconnected.
Air-gapped proofs of concept across government, higher education and defence
Largest model run fully air-gapped at a client site
Shared MCP servers plus one service worker today, and growing
Models reachable through one S.A.L.T CI gateway, where providers are allowed
Federal agency
Sovereign assistant with live government search and a knowledge base, on a LAN-only server.
Public university
University deployment with single sign-on for faculties.
Technical university
Core platform deployed, MOM Studio rolled out with user training, TERUSAN next.
Training centre
Tested on a notebook with no network connection, on-site in a defence environment.
Proofs of concept were delivered with limited features and 5–10 users each. Transfer-of-technology sessions delivered to universities and agencies.
Delivered through partners
Authorised partners
S.A.L.T CI reaches clients through authorised resellers and channel partners across Malaysia, supported by joint meetings, demonstrations and proof-of-concept programmes.
Technology partner
An OEM technology partnership with Dell Technologies pairs enterprise hardware with S.A.L.T CI software for joint go-to-market.
How a proof of concept runs
One real process and the people who run it
On the client's own server, offline if required
With a small group of real users
Prove it keeps working with no internet
Questions worth asking any AI vendor
Six questions that separate a sovereign platform from a hosted service with a local front end, and S.A.L.T CI's answers.
| Question | S.A.L.T CI answer |
|---|---|
| Where do our documents and chats live? | On your servers. Models, knowledge and logs never leave the site unless you allow it. |
| Does it still work with the internet cut? | Yes. Air-gap is a mode of the core, proven at four client sites. |
| Can we put our own brand on it? | Yes. Client and reseller white label, under S.A.L.T CI's own licence. |
| Are we locked to one model or vendor? | No. Switch models freely; open foundations are credited openly. |
| Does it understand local work? | Bahasa Melayu and dialects, halal-aware planning, local payment formats and Malaysian time. |
| What does each question cost? | No per-token fees on local models: unlimited queries at the cost of your own hardware. |
Bring one real process and one server.
We run it there, then disconnect the internet.
Add applications and sites without changing product.
Part of one sovereign platform
Platform runs on the TERAS core with the same sign-in, audit and models as every other S.A.L.T CI application.