Architecture

Deployment topologies for every security posture.

Piixie runs on Mac, Windows, and Linux. Choose the topology that fits your infrastructure: fully offline on a single machine, shared across a team on a local server, connected to a public LLM with local pre-anonymization, or wired into your own cloud LLM. Your data, your rules.

1

Standalone Desktop (Fully Offline)

Zero network. Zero cloud. Everything on your machine.

The simplest and most private deployment option. Piixie Desktop ships with a bundled local LLM that handles all PII detection and anonymization directly on the user's machine.

No internet connection is required at any point in the pipeline -- not for installation, not for model downloads after initial setup, and not for processing documents.

This topology is ideal for air-gapped environments, classified document handling, individual practitioners working with sensitive client data, and any situation where data must never leave the device. The bundled LLM is optimized for PII detection and runs efficiently on consumer hardware, including laptops without dedicated GPUs.

Standalone desktop deployment: a laptop processing documents with a shield icon representing offline privacy

How it works

  1. Load document -- Drag a PDF, DOCX, spreadsheet, image, or any supported file into Piixie Desktop.
  2. Local LLM analysis -- The bundled model scans extracted text for names, addresses, phone numbers, emails, account numbers, and dozens of other PII categories.
  3. Choose anonymization mode -- Redact (black out), replace (placeholder tokens), or synthesize (realistic fake data via local Faker engine).
  4. Export safe copy -- The anonymized document is saved locally. The original is never modified.

Best for

  • Government and defense agencies operating on air-gapped networks
  • Solo practitioners (lawyers, doctors, accountants) handling client files
  • Compliance officers who need to guarantee zero external data transmission
  • Field workers in locations with no reliable internet access
Standalone flow: the user loads a document into Piixie Desktop, the bundled local LLM returns PII locations, Piixie applies Redact, Replace or Synthesize and hands back a safe copy; everything stays on the machine with no network traffic
2

On-Premise Server

Shared team server on your local network. Nothing leaves the building.

For teams that need centralized anonymization without relying on any external service, Piixie Server runs on a dedicated machine within your local network. Workstations submit documents over the LAN, the server processes them through its local LLM, and returns anonymized results.

All traffic stays inside your network perimeter.

This topology is designed for teams of 5 to 500+ who process high volumes of sensitive documents daily. The server handles the computationally intensive LLM inference, which means individual workstations do not need powerful hardware. A single well-provisioned server (with a GPU or high-end CPU) can serve dozens of concurrent users.

On-premise server deployment: a server rack connected to multiple workstations on a local network

How it works

  1. Workstations connect -- Team members access Piixie through a lightweight desktop client or the REST API.
  2. Document submission -- Files are transmitted to the Piixie Server over an encrypted LAN connection. No data routes through the internet.
  3. Server-side processing -- The server runs the LLM, detects PII, and applies the chosen anonymization mode.
  4. Results returned -- Anonymized documents are sent back to the originating workstation.

Best for

  • Hospital systems and healthcare networks processing patient records at scale
  • Law firms with shared document review workflows
  • Financial institutions that require centralized audit logging
  • Enterprise teams that want to offload LLM inference from individual machines

Run it

Start the server on the machine that will host the model — a Linux box with a GPU is the sweet spot. It downloads its model on first run and serves an OpenAI-compatible API on the port you choose:

piixie server --host 0.0.0.0 --port 8787

Pick the larger model if the GPU has the memory:

piixie server --host 0.0.0.0 --port 8787 --model piixie-local-llm-large

On each workstation, add the server as an endpoint in the desktop app (Settings → AI endpoints, OpenAI chat protocol, base URL http://your-server:8787) and pick its model in the model selector. The UI is unchanged — the heavy lifting just happens on the server instead of the laptop. Full setup in the local server guide.

On-premise flow: workstations A, B and C submit documents over the LAN to Piixie Server, which runs batch PII detection on its local LLM and returns an anonymized copy to each workstation; traffic never leaves the local network
3

Public LLM Connection (Safe Proxy)

Anonymize locally first, then safely leverage cloud LLM power.

Sometimes you want the power of a frontier model like GPT-4 or Claude for downstream tasks -- summarization, translation, analysis -- but you cannot send raw PII to a third-party API. This topology solves that problem.

Piixie runs its full anonymization pipeline locally before any data is transmitted, then sends only the sanitized version to the cloud LLM.

The critical guarantee: raw data never reaches the cloud. The cloud LLM only ever sees anonymized content -- blacked-out blocks, placeholder tokens like [PERSON_1], or synthetic fake data. The original PII stays on your machine at all times.

Public LLM connection: documents flow through a local anonymization shield before reaching cloud AI services

How it works

  1. Local PII detection -- Piixie's bundled LLM scans the document and identifies all sensitive entities.
  2. Local anonymization -- PII is redacted, replaced with tokens, or substituted with synthetic data on your machine.
  3. Safe copy created -- A sanitized version is produced containing zero real PII.
  4. Cloud LLM interaction -- The safe copy is sent to OpenAI, Anthropic, or another public LLM API for downstream processing.
  5. Results returned -- The cloud LLM's response references only anonymized entities. Tokens can be re-mapped locally.

Best for

  • Teams that need frontier model capabilities (GPT-4, Claude, Gemini) without PII exposure
  • Research organizations preparing datasets for cloud-based analysis
  • Companies building AI workflows where privacy must be guaranteed at the prompt level
  • Developers integrating Piixie into LLM-powered application pipelines
Public LLM flow: on the user's machine Piixie detects PII with the local LLM and anonymizes it; only the safe copy crosses the network to a cloud LLM such as OpenAI or Anthropic, and the response is re-mapped locally
4

BYO LLM (Bedrock / Azure / Custom)

Full control over the model, the infrastructure, and the data path.

For organizations that have already invested in their own LLM infrastructure, Piixie can connect to your models instead of using the bundled one. This includes AWS Bedrock, Azure OpenAI Service, self-hosted models on dedicated GPU clusters, or any OpenAI-compatible API endpoint.

You control the model, the hardware, and the network path. Your security team approves the model once, your infrastructure team manages the deployment, and Piixie plugs in as a client.

BYO LLM deployment: Piixie connecting to enterprise cloud infrastructure like AWS Bedrock or Azure OpenAI

How it works

  1. Configure LLM endpoint -- Point Piixie at your Bedrock instance, Azure OpenAI deployment, Ollama server, or custom endpoint.
  2. Document processing -- Piixie extracts and pre-processes the document locally.
  3. LLM-assisted detection -- Text fragments are sent to your LLM for PII detection through your infrastructure.
  4. Local anonymization -- All masking, token replacement, and synthetic data generation happens locally.
  5. Export safe copy -- The anonymized document is saved. Audit logs can be sent to your SIEM.

Supported LLM platforms

  • AWS Bedrock -- Claude, Llama, Titan via your AWS account
  • Azure OpenAI -- GPT-4, GPT-4o in your Azure tenant
  • Ollama -- Self-hosted open-weight models on your hardware
  • Custom endpoints -- Any OpenAI-compatible API

Best for

  • Enterprises with existing LLM infrastructure they want to reuse
  • Organizations with strict vendor and model approval processes
  • Teams that need specific models tuned for their domain (medical, legal, financial)
  • Multi-cloud environments where data must stay within a specific cloud provider
BYO LLM flow: Piixie Desktop extracts text locally and sends text fragments to your own LLM (Bedrock, Azure or custom) inside your infrastructure; detected PII entities come back and anonymization happens locally before the document returns to the user
At a Glance

All four topologies, side by side.

Every option keeps the anonymization logic local. The difference is where the LLM inference happens and how data flows between components.

1. Standalone desktop

  1. Your machine
  2. Piixie Desktop
  3. Bundled local LLM
  4. Anonymized output

2. On‑premise server

  1. Team workstations
  2. Piixie Server on your LAN
  3. Server-hosted LLM
  4. Anonymized output

3. Public LLM as a safe proxy

  1. Your machine
  2. Piixie Desktop
  3. Local anonymization
  4. Safe copy only
  5. Cloud LLM OpenAI / Anthropic

4. Bring your own LLM

  1. Your machine
  2. Piixie Desktop
  3. Your LLM Bedrock / Azure / custom
  4. Anonymized output

When to choose Standalone

You need absolute isolation. Air-gapped networks, classified documents, or solo practitioners who want the simplest setup. Install Piixie, process files, done. No server, no accounts, no network.

When to choose On-Premise Server

Your team processes documents at volume and you want centralized control. One powerful server handles the LLM workload so individual workstations stay lightweight. Audit logs, user management, and REST API access come built in.

When to choose Public LLM Connection

You want frontier model capabilities for downstream tasks but cannot send raw PII to the cloud. Piixie anonymizes locally first, so the cloud LLM only ever sees sanitized content. Best of both worlds.

When to choose BYO LLM

Your organization already runs its own LLM infrastructure and you want Piixie to plug into it. Standardize on one model platform, maintain full control over the data path, and satisfy your security team's vendor approval process.

Ready to deploy Piixie in your environment?

Start with the free desktop app. Scale to server mode or connect your own LLM when you need to.

Download Piixie