Private Deployment

For customers whose meeting audio and meeting text may not leave their own infrastructure. Every AI step can run on the customer's machine or in their cloud account:

Step	Default	Private
Transcription (live and post-meeting)	Deepgram	Whisper (STT_PROVIDER=whisper)
Speaker separation	Deepgram	Voice service (/diarize)
Naming speakers by voice	Voice service	Voice service
Summaries, tickets, documents, chat	OpenRouter	Local LLM (LLM_PROVIDER=local)
Document search (embeddings)	OpenRouter	Local model (EMBEDDINGS_PROVIDER=local)
Slide images	OpenRouter (Flux)	Off: slides come without images

The switches are backend-wide environment variables. They are a property of the deployment, not of a workspace.

Starting the stack

From the repository root:

yarn docker:private

This starts Postgres, Redis and Qdrant, plus:

Service	Host port	What it is
whisper	8010	speaches, OpenAI-compatible Whisper server
voice	8002	the voice service (speaker separation, voice profiles), with Sentry off
ollama	11434	local LLM and embeddings server
models-init	none	downloads the Whisper model, the LLM and the embedding model, then exits

The first start downloads the models (about 20 GB with the defaults) into Docker volumes. Later starts reuse them.

On an NVIDIA host, add the GPU override:

cd plan-ai/backend
docker compose -f docker-compose.yml -f docker-compose.gpu.yml --profile private up -d

It switches Whisper to its CUDA image and gives both Whisper and Ollama the GPUs. It needs the NVIDIA driver and the NVIDIA Container Toolkit.

Pointing the backend at it

In plan-ai/backend/.env:

STT_PROVIDER=whisper
WHISPER_BASE_URL=http://localhost:8010

LLM_PROVIDER=local
LOCAL_LLM_BASE_URL=http://localhost:11434/v1
LOCAL_LLM_MODEL=gemma3:27b
LOCAL_LLM_CONTEXT_TOKENS=32768

EMBEDDINGS_PROVIDER=local
LOCAL_EMBEDDINGS_MODEL=bge-m3
LOCAL_EMBEDDINGS_DIMENSION=1024

VOICE_AI_URL=http://localhost:8002

LOCAL_LLM_BASE_URL takes any OpenAI-compatible server, so vLLM, LM Studio or a llama.cpp server work the same way as Ollama.

Choosing the models

LLM. The default is Gemma 3 27B: multilingual (Spanish, Catalan, Arabic, English), about 17 GB in 4-bit, and not a "thinking" model. Models that write out their reasoning before answering can break the strict JSON the pipeline relies on for tickets and summaries, so prefer instruct models. Qwen3 instruct models and Llama 3.3 70B are alternatives when the hardware allows.

The context window matters as much as the model. Ollama defaults to 4,096 tokens and its OpenAI-compatible API can't raise it per request, so the compose file sets OLLAMA_CONTEXT_LENGTH from LOCAL_LLM_CONTEXT_TOKENS. Keep the two in step: the backend uses the value to decide whether to inject a project's documents whole or search them (RAG).

Embeddings. bge-m3 is multilingual and produces 1,024-dimension vectors. The cloud model produces 1,536, so local vectors live in their own Qdrant collection, named after the model and size (for example context_files__bge-m3_1024). Switching back to the cloud finds the original collection untouched. Documents indexed before the switch must be indexed again to be searchable.

Whisper. See Speech-to-Text.

Hardware

Setup	What runs	Rough cost
One GPU with 48 GB (L40S, RTX 6000 Ada)	Gemma 3 27B, Whisper large-v3-turbo and embeddings together	7,000 to 9,000 EUR for the card, or about 1 USD/hour rented
One GPU with 24 GB	Gemma 3 27B alone, Whisper on CPU
CPU only	Whisper and the voice service fine for the post-meeting pass; an LLM of 27B is too slow for interactive chat

These are sizing guides, not measurements on the target hardware. Measure with real meetings before quoting a customer.

Locking the stack down

The compose file publishes every port on 127.0.0.1 only. Docker writes its own firewall rules, so a port published on 0.0.0.0 would be open to the network even behind ufw. Put a TLS proxy in front of the backend and the web app, and keep the other ports closed.

The defaults in the compose file are for development. A private install sets its own values in plan-ai/backend/.env, which compose reads as well:

POSTGRES_PASSWORD=...            # DATABASE_URL must use the same password
REDIS_PASSWORD=...               # and REDIS_URL=redis://:<password>@localhost:6379
QDRANT_API_KEY=...               # also uncomment QDRANT__SERVICE__API_KEY in docker-compose.yml
VOICE_AI_API_KEY=...             # shared by the backend and the voice service
SECRETS_ENCRYPTION_KEY=...       # openssl rand -base64 32, then run yarn secrets:encrypt
API_ADMIN_KEY=                   # empty, or 32+ random characters
CORS_ORIGINS=https://plan.example.com

POSTGRES_PASSWORD only takes effect when the database volume is first created. Qdrant's telemetry is off in the compose file. The queue dashboard stays off unless BULL_BOARD_USER and a BULL_BOARD_PASSWORD of 16 or more characters are set. An internal Jira or Twenty must be listed in SSRF_ALLOWED_HOSTS, because the backend refuses to fetch private addresses. All of these are described in Environment Variables.

What still leaves the machine

Being precise about this is part of the offer:

• Sign-in goes through Firebase Authentication (Google), in the customer's own Firebase project. It sees accounts and emails, not meetings. The web app also loads Google Analytics for Firebase from the same project.
• Recordings and uploaded files are stored in the Firebase Storage bucket set in FIREBASE_STORAGE_BUCKET. For a private deployment, use a bucket in the customer's own Google Cloud project and region.
• Integrations (Jira, Linear, Notion, Twenty, Google Drive…) send what the customer chooses to sync, to the tools the customer connected.
• Error reporting goes to Sentry only when SENTRY_DSN is set; leave it unset. The voice service in this stack has SENTRY_DSN="".
• Email goes through Resend only when RESEND_API_KEY is set. It carries the recipients and the content of the notes a user sends, and invitations.
• Diagram images for the Telegram prospect flow go to https://mermaid.ink unless MERMAID_INK_URL is off or points to a mermaid.ink you host. Nothing else uses it.
• Web searches made by the assistant go to DuckDuckGo with the search query.
• Slide images are off in local mode, because their prompt is written from the meeting's content.

Replacing Firebase with a self-hosted identity provider and object storage (Keycloak and MinIO, for example) is a separate project.

Checking it

# Transcription path, live and post-meeting
yarn --cwd plan-ai/backend stt:smoke --mic sample.wav --language es

# Usage is logged as provider LOCAL at zero cost
psql "$DATABASE_URL" -c "select provider, model, count(*) from \"AiUsageLog\" group by 1, 2;"