Not an app.
An operating system.
SOTA OS · PROGRAMMABLE END TO END
Compare the new SBV circular with our lending policy, cite every clause we need to change, and flag anything that needs my sign-off.
Replace scattered ChatGPT, Claude and Gemini seats with one governed platform, deployed where your rules say, up to fully air-gapped. Then our forward-deployed engineers build the agents each department actually needs.
Book an assessmentAUDIT LOG · 0 ENTRIES
- customer_list.xlsxNOT LOGGED
- payroll_2026.csvNOT LOGGED
- ACME_contract_v4.pdfNOT LOGGED
$2,690 a month across six vendors. No owner, no volume price.
3 files sent to personal accounts. None of it logged.
0 connections to SAP, Salesforce or approvals.
Your teams already use AI.Just not the same one.
Follow the people, the money and the files. One company goes in. Six vendors, three leaks and zero connected systems come out.
You draw the border. The gateway enforces it.
Most enterprises land on hybrid: data stays in your datacenter, and one gateway decides what may reach a cloud model, masking personal data first. Or keep all of it inside.
CLOUD
Live tomorrow
Hosted by us, encrypted at rest, content-free telemetry. Every team on one governed platform in days, no hardware to buy.
HYBRID · RECOMMENDED
Model out, data in
Only two routes out, both through one gate: the model and web search. Personal data masked first, rules per department, every call logged.
SELF-HOST / AIR-GAP
Everything inside
Your datacenter or private cloud, your GPUs, your weights. Air-gapped builds need no internet at all.
Every model, one bill.
Your rules on who uses which.
GPT, Claude, Gemini, Grok, DeepSeek, Qwen, GLM or Kimi behind one dropdown, replacing the seats each team bought on its own. Marketing can get them all; finance only the models that run inside.
One platform, not scattered seats
Every employee reaches every approved model from the same place, with usage you can see and cap per team.
Whitelist per workspace
Admins choose which models each department may run and set its default. Swap a model and prompts, tools and memory stay put.
Or bring the weights home
DeepSeek, Qwen, GLM, Kimi, Llama and Mistral run as open weights inside your own infrastructure. Same dropdown, same audit trail, zero egress.
Not an app to install. An operating system to program.
Every layer is yours to change: the tools an agent can call, the skills it loads, the slash commands people type, the panels they see, the policies admins enforce. One manifest wires it all.
Not an app.
TOOLS & MCP
Every tool an agent can call. Yours to add.
HTTP tools, remote MCP servers, code sandbox, the user's own machine.
SKILLS
Knowledge loaded on demand.
Versioned with the app. Swap a skill, not the platform.
SURFACES
- artifact panel
- admin tab
- /slash command
- inline renderer
- native page
name: lending-complianceversion: 1.4.2tools: - id: policy.compare auth: scoped-jwtskills: - path: ./skills/sbv-circularsslash: - verb: /comparesurfaces: - artifact: policy-diffdeploy: env: [dev, staging, prod]RUNTIME
Multi-agent, memory, queue, guardrails, credits.
Ship it like software
sota dev gives you a tunnel with production-shaped auth. sota deploy runs preflight before anything uploads. Promotion moves the exact same artifact to prod.
Verified at runtime
Signed descriptors, byte-level integrity checks on every entry, immutable packages. Apps mount natively in the shell, never in an iframe.
Depth without forks
Override a tool, add a skill, swap a surface. The platform stays upgradeable because your customization lives in your app, not in our core.
Reliable agents show their work.
No black box. Every step, from parsing a scanned contract to ranking an ERP row, is visible, tunable and logged against the answer it produced.
UseUpdate Opportunityto move the Acme renewal to Negotiation, and note what we agreed on yesterday's call.
Did 4 actions in 9s
- Matching the call to the open deal
- Microsoft Teams Search Messages · “Acme sync”, 3 SepTool
- Salesforce Get Opportunity · Opp-00931Tool
- Salesforce Update Opportunity · Stage, Next stepTool
- Done
Moved Acme renewal (Opp-00931) to Negotiation and set its next step to “Send revised quote by Fri, 12 Sep”, as agreed on the callAcme sync · 3 Sep.
- Stage: Proposal → Negotiation
- Next step: Send revised quote by Fri, 12 Sep
Used 2 sources
Ingest anything
PDF, DOCX, XLSX, images, DWG/DXF/IFC, Drive and SharePoint. Layout-aware OCR that survives 300-page scans.
Trace every answer
Ask in Vietnamese, English or Japanese. Each answer points to the page, row or message. If the source is not there, the agent says so.
A person signs off
For quotes and financial work the agent hands over a working paper: formulas shown, sources linked, assumptions highlighted. Your reviewer approves in minutes.
Built first for the industries that cannot afford to guess.
Each pack is a set of apps, skills and guardrails tuned with a customer in the field. Airlines, banks and hospitals came first because they need sovereignty most.
AIRLINE
Operations that read the manual
Crew, maintenance and ground ops ask one assistant that answers from the approved documents only, with the page cited.
- Manual and bulletin grounding
- Shift handover summaries
- Incident report drafting
BANKING
Compliance that keeps up with the circular
New regulation in, every affected internal policy out, clause by clause, before the deadline hits.
- Regulation impact review
- Policy conflict detection
- Audit-ready obligation register
HEALTHCARE
Records that stay in the building
Clinical and administrative teams work with patient data on hospital infrastructure. Nothing leaves.
- On-prem document intelligence
- Admin paperwork automation
- Role-scoped access, full audit
Live in days. Built for your workflow in weeks.
SotaTek's 1,000+ engineers work as forward-deployed teams: inside your infrastructure, on your real data, until the agent does the job.
- 01
Week 1 · Quick win
Every team moves onto one governed chat with every approved model, on our cloud or yours. The personal subscriptions can go.
- 02
Week 2 · Department survey
We map each team's workflow, data, systems and the accuracy it really needs, then pick the use case that matters most.
- 03
Weeks 2 to 6 · Proof of value
A working agent on your real documents, inside your perimeter, delivered in 14 to 30 days and measured against the success criteria we agreed.
- 04
Then · Scale and hand over
Agent apps for each department. Your team owns the apps, the source and the runbook; we stay on call, not in the way.
Start with one quick win.
A 45-minute session. We look at your data, systems and regulator, then agree on one high-impact workflow and what success looks like.