The new release of the Sovereign AI Ecosystem. We delivered everything planned for the third quarter of 2026: 13 new TWIN:DESK and PROXY:AI features, from meetings turned into tasks to AI models running with no internet access. And this is only the beginning
Meetings and translation, Agentic RAG, verified calculations, AI costs and offline models
Knowledge with a hierarchy of authority, a knowledge graph, AI Tool Studio and compliance while models run
Twin Knowledge Graph, Agent Quality Lab and national infrastructure connectors
Every feature is shown in a video. The code on each card matches its item in the development roadmap
From a recording to tasks, from a foreign language to your own knowledge base
2 features · 2 videosTD-514A live or uploaded recording becomes text with labelled speakers and a timestamp for every statement. Markers, comments and AI analysis lead from the conversation to findings and tasks with an owner and a deadline. Export to PDF, Word, TXT and JSON
TD-515Document translation on local models (CPU or GPU) in a segment editor, with translation memory and a company glossary. The DOCX file comes back with its layout intact, and the finished translation goes to the knowledge base in one click
Answers with sources, reports across projects, numbers calculated rather than guessed
3 features · 5 videosTD-516The agent judges what is missing from the material it has gathered, asks follow-up queries and merges every round into one answer. TWIN:INDEX (the PageIndex mechanism) builds a tree of long documents, so search follows chapters and sections, and diagnostics show every round, every source and the context assessment
TD-517A report across many projects is produced from a published plan (DAG) configured once. Every step shows its status, independent branches run in parallel, and when the data changes the same controlled process simply runs again. When a task needs flexibility, the agent works in an agentic loop: it proposes a plan for approval, uses the knowledge base and tools, and the run history shows every step
TD-518The model interprets the task while a dedicated decimal calculator does the maths. Every result comes with a calculation card listing the arguments, the rounding rule and the engine version, and the engine stops a report from being published if a figure does not match
Less configuration, fewer support questions, less retyping
4 features · 4 videosTD-504You describe in plain language how the assistant should work, and TWIN:DESK proposes the description, system prompt, suggestions and temperature. Every change is shown before and after, you pick only the ones that fit and test the model straight away in a chat
TD-505On first sign-in a short guide shows only the parts of the interface the user can access. Administrators see progress in the user list and restart a guide with one click
TD-513Key stages are saved as named versions, for example “Baseline assumptions” or “Version after review”. Previewing never changes the current content, and restoring creates a new version while keeping the whole history
TD-510The local Twin Desk Agent next to the Windows clock moves meetings from a chosen date range in Outlook Calendar into Asana tasks, with the due date, assignee, meeting time and a link to Outlook. A summary shows what was created, what stayed unchanged and what needs attention, and running the sync again creates no duplicates
The full picture of spending, including tools on employees' computers
3 features · 3 videosTD-511The “AI usage” panel in the Twin Desk Agent shows today's and this month's tokens, the number of calls and the API equivalent for tools such as Codex. History covers the current month and the three before it, and the data feeds the central cost report
PA-519Compares each user's subscription with the API-equivalent forecast and recommends keeping, downgrading or upgrading the plan, and when the work needs an API integration it advises keeping the API. The forecasting method is transparent and “what if” scenarios change nothing without a human decision
PA-523“AI in the company. One picture”: cost forecast, subscriptions, AI usage and pricing coverage on one screen. Plus a scenario budget, items needing attention and an operations section with gateway diagnostics, and at the bottom the next step: confirming the price and comparing subscription options
Air gap without compromise
1 feature · 2 videosPA-525Model Manager starts from your infrastructure: the Agent profile shows the processor, memory and GPU cards. It rates the compatibility of catalogue models, an advisor suggests candidates for a business need, and the deployment plan points to the exact package in Harbor. Model Agent on a server in a network cut off from the internet checks compatibility, starts the model and passes a test via Master
Examples from the sectors we work with. At a demo we show them on your documents
Minutes of a council or committee session with labelled speakers, ready for review in minutes, with findings turned straight into tasks
EU correspondence and documents translated on a local model with the office's own glossary, without sending content outside
A portfolio report in which every amount was computed by the calculator and an auditor can reproduce it from the calculation card
Answers from extensive technical documentation and operating manuals, with a pointer to the right chapter and a check that the source is current
AI models in a network fully isolated from the internet, with a controlled catalogue of versions and a test on every server
Notes from a team briefing with labelled speakers, and procedures in named versions you can always return to
Multilingual product documentation with consistent terminology and search that follows the structure of each document
One picture of AI costs, including tools on employees' computers, with a recommendation on which subscription to keep
In the fourth quarter of 2026 SAIE gets 23 more features: knowledge with a hierarchy of authority, a knowledge graph, a no-code tool studio, regulatory compliance enforced while models run and infrastructure distributed across many nodes
TD-425Answers from the right sources, depending on the user's role and each document's weight and currency
TD-426Relationships between documents, concepts and agent outputs that make every answer easy to justify
TD-427Integrations and interface components built without code, by business users
PA-435Regulatory requirements enforced automatically while models are in use
PA-436Consistent rules on access, permitted models, data and response generation for every team
CD-438A faster start to deployment in the CDF methodology with guided configuration
The 2027/Q1 plan includes Twin Knowledge Graph, Agent Quality Lab, National Infrastructure Connectors and AI Governance Compliance Pack. We will reveal the name of that release soon
Full development roadmap16 videos of TWIN:DESK and PROXY:AI at work, narrated in Polish. Each one opens right on this page
TD-514An interview recording: text with speakers, markers, AI analysis and tasks
TD-515A document from English to Polish: segment editor, translation memory, glossary and DOCX export
TD-516Two search rounds, six queries and diagnostics for every step
TD-516PageIndex: a tree of a long document that the agent searches chapter by chapter and section by section, with context and sources assessed
TD-517A published DAG plan, step statuses and a finished file in the chat
An agent handles an urgent ticket: a plan to approve, steps run in the background and an escalation plan in a Word file
TD-518An order of 3 × PLN 12.50, a calculation card and a check before publishing
TD-504A project status assistant: from a description of needs to a chat test
TD-505A role-aware guide and progress visible to the administrator
TD-513A rollout plan in named versions, restored without losing history
TD-510A week of Outlook calendar meetings moved into Asana tasks
TD-511The “AI usage” panel: tokens, calls and API equivalent from Codex history
PA-519Subscription versus API equivalent and a plan recommendation per user
PA-523The executive panel: forecast, scenario budget and items needing attention
PA-525Infrastructure profile, compatibility rating, model advisor, Harbor packages and a deployment plan
PA-525The agent on a server: resources, model deployment, compatibility check and a test via Master
https://www.allclouds.pl/en/saie-pantera/