Products / TERAS
S.A.L.T CI TERAS
The sovereign AI core. One application that runs your models, gives people a chat workspace, searches your documents, connects tools and agents, and governs all of it, on your own servers, with or without the internet.
Organisation letterhead · 1 pageOpen
AI your organisation can own, trust and grow
Government agencies, universities, defence and regulated enterprises need AI, but cannot send their data to a foreign cloud. Until now, a sovereign setup meant stitching borrowed parts together.
Data leaves the country
Popular AI services run in foreign clouds, under foreign law.
A stack of borrowed parts
A model runner, a chat UI, a vector store and pipelines: four products to install, upgrade and brand.
No proof of quality
Nobody can say which model is fit for which job before people rely on it.
What TERAS changes
| Before | With TERAS |
|---|---|
| A model runner, a chat UI and pipelines from different projects | One product that runs models, chat, knowledge, tools and governance, upgraded together |
| Branding locked by a licence | Your brand, fully: client and reseller white label |
| One custom model per person | Personalisation applied to every model: about me, memory, style |
| Guessing which model is good | Model Studio: playground, test sets, judge, scorecards, fine-tuning |
| Manual copies between servers | Linked sites, multi-node failover and encrypted backups |
| Logs that can be edited | A hash-chained audit trail that shows any change |
TERAS in one line
One sovereign AI core that runs air-gapped on a single notebook and grows, without changing product, into a federated, multi-node data centre.
Every feature works with no internet.
Cloud models only for permitted people.
BM and English in OCR, voice and templates.
Written by S.A.L.T CI, open foundations credited.
Everything an AI platform needs, in one product
People use the Console and the admin panel in a browser or as an app; applications use the /v1 API. Everything a client changes is a setting in the panel, never code.
Run
Model store and engines on CPU, GPU or both, on this server or other machines.
Chat
The Console: answers, attachments, citations, artifacts, voice.
Know
Document collections with OCR, hybrid search and citations.
Act
Tools, MCP, skills, agents, schedules and a code sandbox.
Reach
WhatsApp, Telegram, Teams, Slack, e-mail and a website widget.
Improve
Model Studio: test, compare, grade and fine-tune.
Personalise
About me, memory and style for every person, on every model.
Protect
Sites, nodes, backups, SSO, quotas, guardrails and audit.
A chat workspace that shows its work
Answers stream in, cite the documents they used, and can produce real files. Pages, diagrams and documents open beside the chat.
Any attachment
Documents are read and cited; pictures go to vision models or OCR; audio is transcribed.
Artifact panel
Pages, Mermaid diagrams, SVG, long documents and code, with versions and download.
Real files
Word, PDF, PowerPoint and Excel, on the organisation's own templates.
Talk to it
Speak instead of typing; answers read aloud in English or Malay.
Private when needed
A private chat is never remembered; per-chat knowledge, tools and style.
Everywhere
Phone layout, installable app, seven window styles, personal colours.
Its own model runtime, on any hardware
TERAS runs models itself, with engines started on demand and placed on the GPU, the CPU or both. No separate model server to install. Models are pulled by name from public model libraries, and other TERAS sites and cloud providers are governed like local models.
Any GPU, any age
NVIDIA (CUDA), AMD (ROCm, Vulkan), Intel, Apple (Metal) and older cards through Vulkan; CPU only works too.
Compute agents
Any spare machine runs engines for the site; it keeps only model files.
Custom models and LoRA
Prompt, knowledge, tools and skills on a base; many LoRA adapters share one engine.
Smart routing
Auto model, fallback chains, semantic cache, quotas and cost reports.
Prove a model before people rely on it
Ask once and compare up to three models side by side, each answer with its full trace. Grade them on a test set, then teach a model the organisation's way: build a dataset, train a LoRA adapter on your own GPU, and TERAS tests it before and after.
Import, take from chats, knowledge or a test set
Download the base to train on
On a workstation card or a server
Packed for the TERAS engines
Served beside the base
Before and after, on the same test set
Example: 0% to 100% in minutes
A small model learned a helpdesk answer format from 61 examples: 35 training steps on a workstation GPU, then added to TERAS and scored before and after on its test set.
Any source
OpenAI messages, ShareGPT, Alpaca, chats, test sets, or generated from a knowledge collection.
LoRA or merged
Small adapters served per request, or a merged model.
Small GPUs welcome
Small models train on a workstation card; larger ones on servers.
Answers from your own documents, with sources
Put policies, handbooks and reports in a collection. TERAS reads them, scans included, and every answer points to the passage it used.
Every format
PDF (scans too), Office, OpenDocument, e-mail, e-books, CSV, HTML, pictures.
OCR in BM and English
Text recognition, or a vision model for handwriting and tables.
Hybrid search
Vector plus keyword, re-ranked; sources open at the passage.
Connectors
Folders, SharePoint, OneDrive, Google Drive, Nextcloud, websites and databases.
Ask your databases
Read-only connections: ask in plain language, see the query and the answer.
Official-looking files
Word letterheads, PowerPoint masters and Excel templates kept; titles and dates filled in.
Federated knowledge
Search a linked site's collection where it lives, without copying documents.
Classified collections
Collections carry a security label that decides who may search them and where they may be processed.
From answers to action
Models call tools in a governed loop, agents work toward goals, and assistants meet people where they already are.
Time, calculator, knowledge, web, documents
HTTP or SSE, with per-tool switches
Audio, IRIS, LOGHAT, PERISAI and more; files both ways
Risky operations off by default; every call shown and audited
Agents
A goal with a model, tools and limits. It plans, acts over several steps and reports; changes to other systems wait for approval.
Scheduled tasks
A prompt or an agent every morning or on a date; results to a chat, a file, e-mail or WhatsApp.
TERUSAN flows
TERAS is the reasoning engine of TERUSAN process flows, and agents can start a flow.
Code sandbox
Python in an isolated, network-less container for data, maths and charts.
Deep research
Many steps across knowledge (and the web in hybrid mode) into a cited report.
Skills and filters
Reusable instructions with their own tools; masking, word lists and footers.
Channels
An assistant on WhatsApp Business, same accounts and policies.
Teams, Slack, Telegram
The same assistants as bots in your workspace.
Website widget
A chat for public or intranet sites, tied to chosen knowledge.
E-mail and alerts
Assistants that answer a mailbox; alerts by e-mail and WhatsApp.
Every model answers like it knows you
People tell TERAS who they are once. Their details, memories and preferred style reach every model they use. No more one custom model per person.
About me and style
Name, position, languages and main style, plus styles for situations such as formal BM proposals or concise coding.
Memory
TERAS remembers useful facts after an answer; people can see and forget them.
Never sensitive
Health, religion, politics, race, money, ID numbers and passwords are never remembered (PDPA).
Admins see counts only
Per model: full, facts only or off. Organisation personas for a house style.
Voice in and out
Speech to text and read-aloud voices in English and Malay; local or cloud.
Pictures
Image generation on the server, or cloud models in hybrid mode.
Meeting minutes
A recording becomes a transcript with speakers, then minutes in the organisation's template.
Four languages
Translate between Malay, English, Chinese and Tamil, keeping document layout.
Dialects
Malaysian dialects for speech in and out through LOGHAT; hands-free voice mode.
From a notebook to a national network
A notebook, a department server and a data centre run the same product. Install small, then add nodes, compute agents and linked sites without reinstalling.
One install is one site, fully air-gapped if required
Each node a full copy; failover by itself with three or more
Spare machines lend engines, any GPU or CPU
Independent sites link with a one-time code; requests are signed
What a link can share
| Share | What it does |
|---|---|
| Models | All or chosen models, shown as site/model; cloud models only while the giver is hybrid |
| Engines (burst) | Their engines answer when a model is missing here, when busy here, or both |
| Knowledge | Collections searched where they live; documents never copied |
| Backup space | Holds the other site's encrypted backups, which it cannot read |
| People | Recognise their people, sign them in, and let chats and memories follow a person |
Example: a university that starts small
Lecturers first run TERAS on their own notebooks. Faculty servers join as the main for people without one and as burst for those who have one, backing each other up. Later a campus data centre becomes the main site for compute and data, with every earlier install still linked.
Encrypted backups
Scheduled, AES-256-GCM, to a folder, another site, S3, Google Drive or OneDrive.
No downtime
Drain and remove a node, or hand the main role over, while people keep working.
All in the panel
Connect, accept, join and promote without scripts or the command line.
A few small services acting as one
Only the Gateway is published. Everything else stays inside the deployment: model manager, media, trainer, knowledge worker, automation and channels, on one PostgreSQL database replicated per site.
Compatible with what is already built
| Interface | For |
|---|---|
/v1/chat/completions · /v1/responses | OpenAI-compatible apps and agents, 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 |
/v1/audio · /v1/images | Speech to text, read aloud, pictures |
/mcp · webhooks | TERAS as an MCP server for other agents; signed events for other systems |
Sovereign by design, accountable by default
Every request, sign-in, change and tool call goes into an audit log chained by hash, so any alteration or deletion shows: who, what, which model, where it ran, the result and the tokens.
Data classification routing decides where each request may be processed.
Air-gap enforced
Cloud needs hybrid mode and the person's permission; cloud items are labelled.
Data loss prevention
IC numbers, accounts and secrets masked, warned or blocked per route.
Injection screening
Documents, web pages and tool results checked before reaching the model.
Strong sign-in
Argon2id, authenticator codes, SSO, LDAP, SAML and SCIM.
Signed between sites
HMAC-SHA256 requests, time-limited and used once.
Red-team tested
Moderation and red-team test sets in Model Studio, with PERISAI.
Your people, your rules, your brand
Single sign-on, roles, groups, keys and quotas in one panel, with a view-only admin role for oversight.
Sign-in your way
Google Workspace, Microsoft Entra ID, Keycloak, any OIDC, SAML, LDAP; SCIM provisioning.
Keys with limits
Site and personal keys with permissions, model limits, expiry and rotation.
Quotas and tenants
Limits per person, group or department; several organisations on one site.
Two-level white label
Client branding
The client's name, logo, colours and window styles; S.A.L.T CI co-branding remains.
Reseller white label
The reseller's brand throughout; a small "Built on S.A.L.T CI Framework" notice remains.
Licence and fleet
Each site carries a signed licence: online activation or an offline file for air-gapped sites. An expired licence never locks people out of their data. A fleet view shows versions, health and updates, with signed online and offline updates.
Runs on whatever you already have
Not tied to any machine, vendor or size. Mix new and old cards, vendors, GPU and CPU machines in one site.
| 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 |
Sizing
16–32 GB memory, 8 GB GPU, models up to ~8B.
64–128 GB, one or two 16–48 GB GPUs, up to ~32B.
Multi-node GPU site plus agents; 70B and larger.
Several linked multi-node sites sharing compute.
Install in minutes, upgrade in one click
A Windows installer with a system check, one script for Linux and macOS, a Helm chart, and an offline bundle on USB for sites that never connect. Upgrades take a backup first. Runs on notebooks and rugged machines, bare-metal servers, VMware, Hyper-V, Proxmox, KVM, sovereign or public cloud, Docker, Podman or Kubernetes.
One accountable core instead of three products
TERAS compared with the previous S.A.L.T CI generation and a typical open-source chat stack, as an organisation would install it.
| Capability | TERAS | Previous generation | Open-source chat stack |
|---|---|---|---|
| Runs models without another product | Yes | Needs a model server | Needs a model server |
| Any GPU vendor or age, mixed in one site | Yes | Partly | No |
| OCR in Malay and English built in | Yes | External | External |
| Personalisation for every model | Yes | Model per person | Basic |
| Test sets, judge, scorecards, fine-tuning | Yes | No | Arena only |
| Linked sites and multi-node failover | Yes | Pipeline only | No |
| Hash-chained audit, classification routing | Yes | No | No |
| Encrypted backups to five targets | Yes | No | No |
| Full white label for clients and resellers | Yes | Licence limits | Licence limits |
Built in the open
TERAS is written by S.A.L.T CI. It stands on credited open-source parts such as llama.cpp, whisper.cpp, stable-diffusion.cpp, Piper, Tesseract, PostgreSQL and pgvector, listed in its NOTICE file.
Proven where it matters
S.A.L.T CI ran air-gapped at government, university and defence proofs of concept. TERAS makes that the default.
Part of the S.A.L.T CI platform
TERAS is the brain. Tool servers are the senses: Audio the ear, IRIS the eye, LOGHAT the voice. REKA, PERISAI, TETAMU, TERUSAN, MOM Studio and S.A.L.T NEXUS are the faces.
Part of one sovereign platform
TERAS runs on the TERAS core with the same sign-in, audit and models as every other S.A.L.T CI application.