ATLAS
Sovereign Enterprise AI Execution & Agent Studio
Grounded enterprise knowledge, visual multi-agent orchestration and self-healing code execution inside isolated Linux sandboxes, running on-premises on customer NVIDIA GPUs.
- Answers grounded in your documents, policies and databases, with source citations
- Agent Studio: a visual canvas for governed multi-agent workflows
- Generated code runs only in ephemeral, isolated Alpine Linux sandboxes
- On-premises and State Data Centre deployment with no outbound network egress
Overview
Regulated enterprises want the productivity of AI but cannot hand data control to public clouds. ATLAS moves beyond autocomplete: it grounds responses in internal document repositories, database tables and policies, lets business teams and engineers compose governed agents visually, and executes generated code inside isolated Linux sandboxes that verify the outcome on real data.
ATLAS brings four layers together in one platform — knowledge grounding, an assistant, the Agent Studio and an execution sandbox — so knowledge retrieval directly powers agentic execution without manual copying between tools. It runs on customer NVIDIA GPUs with no external cloud dependencies.
Key capabilities
-
Multi-format document ingestion
Ingests PDFs, Word documents, OpenAPI specifications, policy manuals and database schemas into on-premises vector embeddings.
-
Citations and verification
Assertions are tied to the exact source chunks, with page and line citations, so every answer can be traced to its evidence.
-
Live database schema grounding
Connects to operational SQL and NoSQL stores and translates business questions into validated, read-only queries.
-
Visual Agent Canvas
Design, configure and chain multi-agent systems from idea to production without writing boilerplate code: drag-and-drop trigger and decision nodes, role-based task routing.
-
Pre-deploy test sandbox
Simulate and evaluate agent execution against real and synthetic data in a safe preview environment before exposing agents to users.
-
Human-in-the-loop governance
Mandatory approval checkpoints, four-eyes sign-off for critical actions and natural-language guardrails keep agents within enterprise policy.
-
Isolated execution sandboxes
Generated code never runs on the host: it executes in isolated containers with no capabilities, no host mounts and process, CPU and memory limits — optionally with no network at all.
-
Self-healing execution
When a script fails (missing library, syntax mismatch, database constraint), the engine feeds stderr back to the agent to patch the code automatically over multiple turns.
-
Deterministic failure mode
If execution does not pass verification within its step budget, the task ends as FAILED; unverified code is not presented as complete.
-
Parametric UI for business users
Non-technical users work through safe forms, charts and report widgets instead of Python or Bash.
-
Native MCP Forge
Uses the open Model Context Protocol so local open-weight models can discover, authenticate to and query enterprise systems through pre-built connectors.
-
Automated quality and reasoning inspection
Automated output quality scoring and a step-by-step reasoning inspector for agent runs.
-
Research notebooks
Workspaces built from chosen sources: guides, briefing documents, study guides, FAQs, timelines, mind maps and two-host audio overviews, all grounded with citations.
-
Knowledge graph
Entities and relationships extracted from every document into a permission-aware graph explorer with source passages.
-
AI gateway and governance
Policy-based model routing, budgets, a kill switch, hash-chained audit of every call, and guardrails for prompt injection and Indian personal data.
Architecture
ATLAS is organised as four converged layers on a single on-premises platform. Grounded knowledge feeds the assistant and the Agent Studio, and every action that needs code runs in the sandbox layer under guardrails and approval gates.
-
Layer 1 · Grounding
- Multi-document RAG indexing
- Policy and database grounding
- Automated report synthesis
-
Layer 2 · Assistant
- Full codebase reasoning
- Business process workflows
- Multi-department productivity
-
Layer 3 · Agent Studio
- Visual workflow canvas
- Multi-agent chain builder
- Human-in-the-loop checks
-
Layer 4 · Sandbox
- Ephemeral Alpine Linux containers
- Self-healing error loops
- Parametric UI
Technical specifications
| Deployment unit | Docker Compose stack of 33 containers on a private network; only nginx (80/443) exposed, all internal ports bound to 127.0.0.1 |
|---|---|
| Services | Go control plane (API gateway, auth, RBAC, identity/SSO, documents, search, admin, licensing, monitoring) and Python AI services (RAG engine, embeddings, OCR, LLM router, knowledge graph, voice, model manager) |
| Front end | Next.js user portal and admin console |
| Stores | PostgreSQL 16 (system of record), Redis 7, MinIO object storage, Weaviate (vector + BM25), Memgraph knowledge graph with optional read replicas |
|---|
| Document formats | PDF (incl. scans), Word, Excel, PowerPoint (incl. legacy .doc/.xls/.ppt, OpenDocument), RTF, CSV, HTML, text, Markdown, images, e-mail (.eml/.msg) |
|---|---|
| OCR and languages | Layout-aware extraction; OCR in English and 13 Indian scripts with automatic script detection |
| Search | Hybrid (keyword + semantic) search with document ACLs; multilingual embeddings that work offline (default intfloat/multilingual-e5-small) |
| Citations | Every answer cites passages by document, page and paragraph |
| Local runtimes | Ollama, vLLM, LM Studio |
|---|---|
| Optional cloud providers | OpenAI, Azure OpenAI, Google, Anthropic, Hugging Face — with PII masking before any cloud call |
| Routing | By priority, latency, cost, quality, load, failover or custom rules; per-model token and query limits |
| Ask a database | PostgreSQL, MySQL, MongoDB, Cassandra, Elasticsearch — read-only queries with time and row limits |
|---|
| Connectors | Model Context Protocol (Streamable HTTP and HTTP+SSE); Google Drive, OneDrive, SharePoint, Dropbox, Box, e-mail; S3, MinIO, SFTP, NFS/SMB |
|---|
| Sign-in | Local users and SSO (Azure AD, Okta, Google, Keycloak, OIDC, SAML, LDAP) with directory sync; TOTP two-step sign-in with recovery codes |
|---|
| Speech | Offline speech-to-text (Whisper) and text-to-speech (Kokoro); Indian-language speech output |
|---|
| Isolation | Code sandboxes with no capabilities, no host mounts, process/CPU/memory limits; optional no-network sandboxes |
|---|---|
| Air-gap mode | AIRGAP_MODE=true disables all outbound calls from chat |
Deployment
-
On-premises / sovereign
Runs on your own servers with local models and embeddings.
-
Air-gapped
Local models only; outbound calls from chat disabled and code sandboxes without network.
-
Hybrid
Sensitive workloads to local models, others to approved cloud providers, with PII masking on everything sent out.
-
GPU or CPU
NVIDIA GPUs used automatically when present; CPU-only installs supported.
Requirements
| Operating system | Ubuntu 20.04, 22.04 or 24.04 (x86_64); 22.04 or 24.04 recommended |
|---|---|
| CPU / RAM | Minimum 4 cores / 8 GB; recommended 8+ cores / 16–32 GB |
| Disk | Minimum 40 GB free; recommended 100 GB+ SSD |
| GPU | Optional; NVIDIA GPU with 16 GB+ VRAM recommended for local models (CPU-only installs supported) |
| Installer | setup.sh installs the NVIDIA driver (if needed), Docker, NVIDIA Container Toolkit, generated secrets and a systemd service |
| Rollout plan | Days 1–7: air-gapped sandbox proof of concept · Days 8–20: RAG and MCP tool indexing · Days 21–30: production rollout Source-reported |
Supported integrations
-
PostgreSQL, MySQL, MongoDB, Cassandra, Elasticsearch
Databases
Plain-language questions become read-only queries with limits.
-
Model Context Protocol (MCP)
Connected systems
Connect any MCP server or build tools in MCP Forge; tools that change data stay off until approved.
-
Google Drive, OneDrive, SharePoint, Dropbox, Box, e-mail
Content sources
Connected libraries searchable in chat.
-
S3, MinIO, SFTP, NFS/SMB
Storage ingestion
Scheduled ingestion with folder mappings.
-
Azure AD, Okta, Google, Keycloak, OIDC, SAML, LDAP
Identity
SSO and directory sync.
-
Ollama, vLLM, LM Studio; OpenAI, Azure OpenAI, Google, Anthropic
Models
Local and optional cloud models behind one router.
Security & governance
-
Process and network isolation
Generated code executes only in ephemeral sandboxes with no outbound network access.
-
Guardrail interception
Code is scanned before execution for unauthorised network calls, embedded credentials and privileged commands.
-
Credential sealing
AES-GCM sealed secrets; protects against credential exfiltration via prompt injection.
-
Destructive-query prevention
Schema-aware validation blocks destructive SQL commands.
-
Approval gates
Human-in-the-loop checkpoints and four-eyes sign-off for critical actions.
-
Execution telemetry
Complete telemetry for every sandbox execution.
Use cases
-
BFSI
Automated underwriting and SQL audit
Ingests financial dossiers, synthesises risk models and runs automated balance reconciliations inside sandboxes with no cloud egress.
-
Government / State Data Centres
Multi-agent grievance workflows
Chained triage, escalation and database-update agents, composed visually, resolve citizen grievances inside State Data Centres.
Documentation & developer resources
Architecture
Security
Deployment
Downloads
- ATLAS — Technical & Customer Deck
ATLAS technical and customer briefing (PowerPoint, 10 slides).
Frequently asked questions
Does ATLAS need a public cloud?
No. ATLAS is designed to run on customer NVIDIA GPUs on-premises or in a State Data Centre, without external cloud dependencies.
What happens when generated code fails?
The engine feeds the error output back to the agent to repair the code over several turns. If verification still fails within the step budget, the task is reported as FAILED rather than complete.
How are tools and data sources connected?
Through the open Model Context Protocol (MCP) and pre-built connectors for databases, internal REST APIs and Kafka streams. Credentials are sealed and never exposed to the model.
Items marked Source-reported are as stated in VedaCLOUD.AI product material and have not been independently verified.