PROXY:AI - the AI gateway that controls models, data and costs

An AI gateway is a single entry point between users, applications and AI agents and the models: it checks every request, protects data, selects the model and records an audit trail. PROXY:AI combines an AI gateway and an AI firewall in one central control layer for over 100 models - cloud and local. It eliminates shadow AI, protects sensitive data and keeps costs under control - from cloud to on-premises and air-gap deployments

Digital sovereignty: one day without your AI vendorCan your office work one day without its AI vendor?
<50 msreal-time analysis
100+supported models
12scanning languages
6+ monthsaudit history on demand
PROBLEM

Shadow AI means three risks at once

Securitysensitive data reaches external models outside policy → AI Firewall masks data and blocks attacks on the model
Costsan agent stuck in a loop generates runaway bills (denial-of-wallet) → Cost Control: budgets and limits
Provider dependencyvendor lock-in → Zero-Code Switch: change models without rewriting applications
Regulatory pressureAI Act, GDPR, DORA, NIS2/Polish KSC, KRI, ISO/IEC 42001 → Governance: WORM records and evidence packs
Mini game · 60 seconds
How much data can you protect in 60 seconds?PROXY:AI [Data Guard - combat mode]

In the game you are the PROXY:AI gateway: protect sensitive data, block attacks and let ordinary work through. The game and the learning material are in the Knowledge section

Play in KnowledgeNo sign-up · PL and EN
One request lifecycle - the same for a person, application or agent
1 · Authentication · RBAC: roles, groups, keys · - no direct API keys · AI GATEWAY · 2 · Context assessment · ABAC: who, from where, which model, · and at what risk · AI GATEWAY · 3 · Input scanning · data masking, blocking · prompt injection and jailbreak · AI FIREWALL · 4 · Routing and budget · model selection based on scan · results and limit checks · COST CONTROL · 5 · Model invocation · or a response from the · semantic cache for · repeated questions · AI GATEWAY · 6 · Output scanning · model response analysis · AI FIREWALL · 7 · Human approval · high-risk operations · (human-in-the-loop) · GOVERNANCE · 8 · Audit record · immutable WORM register · GOVERNANCE · REQUEST: PERSON, APPLICATION OR AGENT · RESPONSE AND AUDIT TRAIL
Zero-Code Switch - manage access profiles, not models
User · sends a request · Application · sends a request · AI agent · sends a request · Assigned profile · access · instead of selecting a raw model · Decision · system · permissions · cost · data class · Local models · e.g. Bielik, Llama · Cloud models · selected by policy · WHO IS ASKING · PROFILE · DECISION · MODEL · over 100 models for policy-based selection
Platform modules

Nine modules across four pillars

A dedicated module implements each stage of the lifecycle. Changing a model or adding a policy or limit does not require rewriting applications

Modules around one gateway
Model Registry · model and access profile catalogue · ABAC Gateway · attribute-based request evaluation · RBAC and identity · roles, groups, keys; rotation without downtime · MCP Gateway · oversight of autonomous AI agents · AI Firewall · scanning in 12 languages, data masking · Guardrails · prompt injection and jailbreak filters · Prompt management · response policy repository · Immutable audit · WORM register, Prometheus and Grafana · FinOps and Anti-DoW · budgets, Twin:Coin, Basic / Advanced classes · PROXY:AI · gateway for services, rules and compliance · AI Gateway · AI Firewall · Governance · Cost Control · AI GATEWAY · AI FIREWALL · GOVERNANCE · COST CONTROL
Regulatory compliance and evidence

A record that cannot be denied

The WORM register stores every model invocation, its cost and decision context immutably and with cryptographic protection. Using data collected during normal operations, PROXY:AI prepares audit evidence packs with one click

Evidence supports demonstrating compliance - it does not replace legal advice or certification
2026-09-23 10:41:07 WORM user=a.nowak · profile=finance-confidential2026-09-23 10:41:07 firewall: national ID ×2 → masking · 38 ms2026-09-23 10:41:08 routing: model=local-bielik · cost=0.42 TC2026-09-23 10:41:09 hash=7f3a…c91e · signature ✓
Prepare an evidence pack
Documentation under AI Act Annex IV✓
DORA risk reports✓
NIS2 / Polish KSC materials✓
ISO/IEC 42001 audit packs
FEATURE CATALOGUE

57 PROXY:AI features

The complete gateway feature list: model access, policies and permissions, compliance and oversight, administration and the administrator panel. Including 3 new in SAIE Pantera

Administration and governance8
  1. AI Cost Forecast & Optimizer: an API-equivalent forecast per user, a subscription plan recommendation and “what if” scenarios that change nothing without a human decision
  2. A layer for knowledge billing, monitoring and budgeting of AI resource use
  3. Logging of operational events, administrator analytics panels and monitoring of provider errors
  4. Defining rules on which AI actions are permitted, which require approval and which tools are blocked
  5. A confidentiality mode for working with sensitive data
  6. An MCP gateway with multi-level control of access to tools at user, role and team level
  7. Restricting and auditing access to tools, functions, code execution and the terminal
  8. Configuring transformations of AI requests and responses without writing code, using simple administrative rules
Compliance and model oversight8
  1. Controlled use of AI through audit, access policies, human oversight, recording of decisions and documenting the use of models
  2. Readiness for AI Act requirements, including presentation of models, policy decisions, events, risks and the audit trail
  3. Managing the exposure of personal data by detecting, masking, blocking or escalating queries
  4. Registers and operational evidence useful for managing AI in line with ISO/IEC 42001
  5. Regulatory compliance templates for the AI Act
  6. Model SBOM and reporting of compliance with Cyber Resilience Act requirements
  7. Human Oversight mechanisms covering labelling of AI content, FRIA/ASIA export and approval gates for critical actions
  8. A compliance panel presenting policies, access decisions, events, risks and active models
Access to LLM models19
  1. Offline Model Registry: Model Manager matches models to the infrastructure and prepares Harbor packages, and Model Agent installs and tests them in a network without internet access (air gap)
  2. A single point for managing communication with LLM models for applications, agents and users
  3. Support for multiple AI model providers: cloud, local and on-premises
  4. Mapping errors from different providers to a consistent format compatible with the OpenAI API
  5. Model routing according to configuration, availability, cost, policies and user permissions
  6. Visualisation of a query’s traffic route through rules, models, providers and policy decisions
  7. Configuration of fallback models and automatic switchover when the primary model is unavailable
  8. A model catalogue with metadata, including alias, provider, type, status and mode of use
  9. SLA isolation per tier with dedicated instances and GPU hardware isolation
  10. Declarative regional data residency in an EU-only model per tenant
  11. Defining AI resources as Custom Resource Definitions in Kubernetes
  12. Protection against excessive spending by detecting unusual consumption patterns and cutting off suspicious sessions
  13. A dry-run panel for testing ABAC security rules before they go live in production
  14. No-code transformations of AI requests and responses without the need to write code
  15. IDE integration through a dedicated plugin with real-time visibility of costs and tokens
  16. LLM Playground as a test environment for experiments, comparing answers and tuning parameters
  17. A Python and WASM plugin framework that can be extended through a marketplace
  18. Validation of models before they are launched in the AI environment
  19. An internal currency model for settling the use of AI, including per-action costs, budgets and transfers
Access policies and permissions7
  1. Model access decisions based on attributes of the user, group, model, data type and environment
  2. Predefined sector profiles of security policies for regulated industries, including finance, healthcare, public administration and energy
  3. Real-time detection of abuse attempts, including injection, jailbreak, prohibited content and anomalies in token usage
  4. Measuring the cost of every query by user, team, project and model, with hierarchical budgets
  5. A central, versioned prompt registry
  6. Support for prompt templates and for decorating queries with the central gateway’s rules
  7. Caching of repeated queries using semantic similarity, with key isolation per tenant
Administrator panel15
  1. Executive AI Dashboard: cost forecast, subscriptions, AI usage, scenario budget and items needing attention on one screen for the board
  2. A dashboard showing the status of database integrations, outbox/reconciliation, and costs and usage
  3. Knowledge Chat, a conversational knowledge assistant that supports the operator or administrator
  4. A model registry with test statuses, BYOK tests and a routing projection
  5. A system for logging and auditing AI operations, with filters and CSV/JSON export
  6. Unified telemetry with the ability to replay historical requests to AI
  7. Management of virtual keys, covering creation, synchronisation, masking, rotation and audit
  8. Administration of ABAC rules with YAML import and export, a dry-run mode and dispatch to the gateway
  9. A dashboard of ABAC decisions, a HITL queue, user logs and a risk register
  10. AI security control through guardrails, control policies, human review and evidence logs
  11. Handling of operational schedules, Langflow agents and a list of agents that can be launched
  12. Spending and token logs for reporting costs and configuration
  13. Browsing Docker logs with a time filter
  14. Orchestration of complex, multi-step agent processes that keep their state between steps
  15. Automatic or semi-automatic evaluation of models and adaptive real-time A/B testing of prompts
DEVELOPMENT ROADMAP

What is coming to PROXY:AI in the next quarters

32 roadmap items through Q4 2027. The 2026/Q3 features already work in SAIE Pantera

Full SAIE development roadmap

2026/Q4IN PROGRESSSAIE Pantera extension

This quarter’s extension adds features to TWIN:DESK.

2027/Q1SAIE TYGRYS

PA-431Enterprise RAG Gateway

Agents and applications built outside the core product need secure, consistent access to search across corporate knowledge

PA-432Priority OCR Service

High volumes of scanned documents create queues where urgent files compete with lower-priority jobs and processing status is difficult to track

PA-433Session Intelligence Layer

Without recognizing the context of an individual session, it is difficult to apply the right limits, policies, and behavioral analysis to a specific conversation

PA-434Multi-Node Routing Fabric

In a multi-node environment, requests must reach available resources while taking workload, location, and model requirements into account

PA-435AI Compliance Engine

The organization needs to enforce regulatory requirements automatically during model use instead of relying solely on manual compliance procedures

PA-436AI Policy Control

Different teams and use cases require consistent rules for access, permitted models, data, and response generation

PA-437AI Cluster Manager

Distributed AI infrastructure requires a single place to control nodes, capacity, availability, and workload allocation

Questions about PROXY:AI

What is PROXY:AI, and why should IT deploy it?

It is the central control layer - one router for different AI models. It addresses shadow AI and protects the organisation against sensitive data leaks and uncontrolled costs. We show the gateway live in course 11 IT and DPO, together with twenty security requirements for the AI environment. We show the gateway live in course 11 IT and DPO, together with twenty security requirements for the AI environment

ProblemCourse 11 IT and DPO

How does PROXY:AI protect budgets against deliberate cost inflation?

FinOps and Anti-DoW mechanisms set tenant and user limits, including protection against denial-of-wallet attacks

Platform modules

What about high-risk operations?

PROXY:AI supports human-in-the-loop oversight: a critical process pauses until the appropriate person authorises it

Request lifecycle

How can we show an auditor why the system used a particular model?

The immutable WORM audit log records the complete operational history: model, configuration, invocation cost and data needed to reconstruct the decision. In a bank and at an insurer, this log is evidence for KNF supervision and the basis for ICT incident analysis under DORA

WORM register

How does PROXY:AI control AI agents' permissions?

The MCP Gateway lets agents use only explicitly assigned tools, such as Jira, and records an audit trail of their actions

Platform modules

Does PROXY:AI also check what models generate?

Yes. Model responses are scanned again for sensitive data and dangerous instructions

Request lifecycle

Can new security rules be tested without risk?

Yes. The dry-run panel allows ABAC rules to be tested on simulated traffic before production deployment

Feature catalogue

Will the system block personal data from being sent to the cloud?

Yes. The AI Firewall analyses data in under 50 milliseconds and neutralises items such as Polish national ID and tax numbers and email addresses before they leave the organisation's network

Request lifecycle

How does PROXY:AI reduce query costs?

Semantic caching serves repeated questions with similar meanings from the cache, without querying an expensive model again

Platform modules

How does PROXY:AI protect models against user attacks?

Real-time guardrails filter requests and block attacks such as prompt injection and jailbreaks

Request lifecycle

What if we want to move to a newer model?

Zero-Code Switch: change the primary model in PROXY:AI settings without rewriting business application code

Zero-Code Switch

How does the ABAC Gateway differ from ordinary API access?

It does not admit requests blindly. For every request, it evaluates who is asking, from which environment, using which data and at what risk - then allows, restricts or blocks the operation

Platform modules

How is PROXY:AI different from Azure API Management, Cloudflare AI Gateway and LiteLLM?

Azure API Management and Cloudflare AI Gateway are gateways tied to a single vendor’s platform, and LiteLLM is open-source routing software that the organisation extends and maintains itself. PROXY:AI runs in the cloud, on-premises and air-gapped, and beyond routing it adds an AI firewall with data masking in 12 languages, ABAC evaluation, an MCP gateway for agents, a WORM register with evidence packs and budgets

Modules

Does PROXY:AI work with Bielik and PLLuM?

Yes. PROXY:AI supports local models, including Bielik and PLLuM, alongside cloud models. A policy decides which model is used - for example, requests containing confidential data go only to a local model

Zero-Code SwitchBielik and PLLuM

Can PROXY:AI tell us which AI subscription pays off?

Yes. AI Cost Forecast & Optimizer compares 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 Executive AI Dashboard shows the board the forecast, the scenario budget and items needing attention on one screen

Video: AI Cost Forecast

How do we run a new AI model in a network with no internet access?

With the Offline Model Registry. Model Manager starts from your infrastructure profile (processor, memory, GPU cards), 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

Video: Model Manager

How do you keep data secure in AI?

Data security in AI means control over what data reaches models, which models it reaches and who can check this afterwards. PROXY:AI provides it through three mechanisms: it detects and masks personal data before a query leaves the organisation; it records every query in a WORM register that cannot be changed; and it allows only models that meet the policies, for example those running locally or in the EU, and can send queries with confidential data only to a local model. In organisations covered by the amended KSC Act, the query log helps show oversight of AI tools

264 All questions · in 9 categories
NEXT

What else to explore

You are here:SAIETWIN:DESKPROXY:AICDF methodologyDeployment and securityDevelopment roadmap