Project David Platform

projectdavid-platform is a self-hosted AI runtime that implements the OpenAI Assistants API specification, deployable on your own infrastructure, fully air-gapped if required. You get a production-ready API server out of the box: assistants, autonomous agents, RAG pipelines, sandboxed code execution, and multi-turn conversation, all through a single, standards-compliant REST API with full parity to OpenAI. Connect any model, anywhere. Run inference locally via Ollama or vLLM, route to remote providers like Together AI, or span both, all through one unified API surface. Switch providers without changing a line of application code. If your stack already speaks the OpenAI Assistants API, it already speaks Project David.

Your models. Your data. Your infrastructure. Zero lock-in.

---

---

Documentation

Topic	Link
Full Documentation	docs.projectdavid.co.uk
Platform Overview	docs.projectdavid.co.uk/docs/platform-overview
Configuration Reference	docs.projectdavid.co.uk/docs/platform-configuration
Upgrading	docs.projectdavid.co.uk/docs/platform-upgrading
CLI Reference	docs.projectdavid.co.uk/docs/projectdavid-platform-commands
SDK Quick Start	docs.projectdavid.co.uk/docs/sdk-quick-start
Sovereign Forge	docs.projectdavid.co.uk/docs/1_sovereign-forge-cluster

---

Installation

pip install projectdavid-platform

No repository clone required. The compose files and configuration templates are bundled with the package.

Windows users: pip installs the pdavid command to a Scripts directory that is not on PATH by default. If pdavid is not found after installation, add the following to your PATH:

C:\Users\<your-username>\AppData\Roaming\Python\Python3XX\Scripts

Replace Python3XX with your Python version (e.g. Python313). On Linux and macOS this is handled automatically.

---

Quick Start

This section walks you from a fresh install to your first streaming inference response.

1. Start the stack

pdavid --mode up

On first run this will generate a .env file with unique cryptographically secure secrets, prompt for optional values, pull all required Docker images, and start the full stack in detached mode.

To start with local GPU inference via Ollama:

pdavid --mode up --ollama

Requires an NVIDIA GPU with the NVIDIA Container Toolkit installed.

To start the full Sovereign Forge training and inference mesh:

pdavid --mode up --training

This starts the training pipeline, Ray cluster, and the Ray Serve inference worker. See Sovereign Forge below.

---

2. Bootstrap the admin user

pdavid bootstrap-admin

Expected output:

================================================================
  ✓  Admin API Key Generated
================================================================
  Email   : admin@example.com
  User ID : user_abc123...
  Prefix  : ad_abc12
----------------------------------------------------------------
  API KEY : ad_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
----------------------------------------------------------------
  This key will NOT be shown again.
================================================================

⚠️ Store this key immediately. It is shown exactly once and cannot be recovered.

---

3. Create a user and API key

The admin key provisions users. Each user gets their own API key for SDK operations.

import os
from projectdavid import Entity
from dotenv import load_dotenv
load_dotenv()

client = Entity(
    base_url=os.getenv("PROJECT_DAVID_PLATFORM_BASE_URL"),
    api_key=os.getenv("PROJECT_DAVID_PLATFORM_ADMIN_KEY"),
)

# Create a user
user = client.users.create_user(
    name="Sam Flynn",
    email="sam@encom.com",
)
print(user.id)

# Create an API key for that user
api_key = client.keys.create_key(user_id=user.id)
print(api_key)

Store user.id and the printed API key — you will need both for SDK operations.

---

4. Run your first inference

Install the SDK:

pip install projectdavid

⚠️ projectdavid is the developer SDK. projectdavid-platform is the deployment orchestrator. Do not confuse the two.

import os
from projectdavid import Entity, ContentEvent, ReasoningEvent
from dotenv import load_dotenv
load_dotenv()

# Use the user key — not the admin key — for application operations.
client = Entity(
    base_url=os.getenv("PROJECT_DAVID_PLATFORM_BASE_URL"),
    api_key=os.getenv("PROJECT_DAVID_PLATFORM_USER_KEY"),
)

# Create an assistant
assistant = client.assistants.create_assistant(
    name="Test Assistant",
    model="DeepSeek-V3",
    instructions="You are a helpful AI assistant named Nexa.",
    tools=[
        {"type": "web_search"},
    ],
)

# Create a thread — threads maintain the full message state between turns
thread = client.threads.create_thread()

# Add a message to the thread
message = client.messages.create_message(
    thread_id=thread.id,
    assistant_id=assistant.id,
    content="Find me a positive news story from today.",
)

# Create a run
run = client.runs.create_run(
    assistant_id=assistant.id,
    thread_id=thread.id,
)

# Set up the inference stream — bring your own provider API key
stream = client.synchronous_inference_stream
stream.setup(
    thread_id=thread.id,
    assistant_id=assistant.id,
    message_id=message.id,
    run_id=run.id,
    api_key=os.getenv("HYPERBOLIC_API_KEY"),  # or TOGETHER_API_KEY etc.
)

# Stream the response
for event in stream.stream_events(model="hyperbolic/deepseek-ai/DeepSeek-V3"):
    if isinstance(event, ReasoningEvent):
        print(event.content, end="", flush=True)
    elif isinstance(event, ContentEvent):
        print(event.content, end="", flush=True)

See the complete SDK reference at docs.projectdavid.co.uk/docs/sdk-quick-start.

---

Your Architecture

Do not use the platform API as your application backend directly. The intended design is a three-tier architecture:

• projectdavid-platform — inference orchestrator (this package)
• Your backend — business logic, auth, data
• Your frontend — user interface

See the reference backend and reference frontend for starting points.

---

Stack

Service	Image	Description
api	thanosprime/projectdavid-core-api	FastAPI backend exposing assistant and inference endpoints
sandbox	thanosprime/projectdavid-core-sandbox	Secure code execution environment
db	mysql:8.0	Relational persistence
qdrant	qdrant/qdrant	Vector database for embeddings and RAG
redis	redis:7	Cache and message broker
searxng	searxng/searxng	Self-hosted web search
browser	browserless/chromium	Headless browser for web agent tooling
otel-collector	otel/opentelemetry-collector-contrib	Telemetry collection
jaeger	jaegertracing/all-in-one	Distributed tracing UI
samba	dperson/samba	File sharing for uploaded documents
nginx	nginx:alpine	Reverse proxy — single public entry point on port 80
ollama	ollama/ollama	Local LLM inference (opt-in, --ollama)
inference-worker	thanosprime/projectdavid-core-inference-worker	Ray HEAD node + Ray Serve inference mesh (opt-in, --training)
training-worker	thanosprime/projectdavid-core-training-worker	Fine-tuning job runner (opt-in, --training)
training-api	thanosprime/projectdavid-core-training-api	Fine-tuning REST API (opt-in, --training)

---

System Requirements

Resource	Minimum	Notes
CPU	4 cores	8+ recommended
RAM	16GB	32GB+ if running the inference mesh
Disk	50GB free	SSD recommended
GPU	—	Nvidia 8GB+ VRAM, optional, required only for Ollama / Sovereign Forge

Runtime dependencies: Docker Engine 24+, Docker Compose v2+, Python 3.9+. nvidia-container-toolkit required only for GPU services.

---

Lifecycle Commands

Action	Command
Start the stack	pdavid --mode up
Start with Ollama	pdavid --mode up --ollama
Start Sovereign Forge	pdavid --mode up --training
Start Sovereign Forge + Ollama	pdavid --mode up --training --ollama
Pull latest images	pdavid --mode up --pull
Stop the stack	pdavid --mode down_only
Stop and remove all volumes	pdavid --mode down_only --clear-volumes
Force recreate containers	pdavid --mode up --force-recreate
Stream logs	pdavid --mode logs --follow
Add a GPU worker node	pdavid worker --join <head-node-ip>
Destroy all stack data	pdavid --nuke

Full CLI reference at docs.projectdavid.co.uk/docs/projectdavid-platform-commands.

---

Configuration

pdavid configure --set HF_TOKEN=hf_abc123
pdavid configure --set TRAINING_PROFILE=standard
pdavid configure --interactive

Rotating MYSQL_PASSWORD, MYSQL_ROOT_PASSWORD, or SMBCLIENT_PASSWORD on a live stack requires a full down and volume clear. The CLI will warn you.

Full configuration reference at docs.projectdavid.co.uk/docs/platform-configuration.

---

Upgrading

pip install --upgrade projectdavid-platform
pdavid --mode up --pull

After upgrading, pdavid will print a notice on the next run pointing to the changelog. Running --pull fetches the latest container images. Your data and secrets are not affected.

Full upgrade guide at docs.projectdavid.co.uk/docs/platform-upgrading.

---

Sovereign Forge — Private Training + Inference Mesh

Project David includes an opt-in fine-tuning and inference cluster built on Ray Serve. Point it at any NVIDIA GPU — a laptop, a workstation, a gaming rig, or an H100 rack — and it handles training job scheduling, model deployment, and inference routing across all of them simultaneously. Your data and models never leave your machines.

pdavid --mode up --training

This starts three services under a Docker Compose profile:

• inference-worker — Ray HEAD node. Owns the GPU on this machine. Runs Ray Serve and hosts the InferenceReconciler. All vLLM inference is managed here. The main API's VLLM_BASE_URL points to this container.
• training-worker — Fine-tuning job runner. Manages the training job lifecycle via Redis queue.
• training-api — REST API for datasets, training jobs, and the model registry.

Scale-out — adding a second GPU machine

Run this on machine 2, 3, or N. No compose files or full stack installation needed on worker machines — just Docker and the NVIDIA Container Toolkit.

pip install projectdavid-platform
pdavid worker --join <head-node-ip>

Ray discovers the new node automatically and the InferenceReconciler distributes load across all available GPUs.

Full documentation at docs.projectdavid.co.uk/docs/1_sovereign-forge-cluster.

---

Docker Images

• thanosprime/projectdavid-core-api
• thanosprime/projectdavid-core-sandbox
• thanosprime/projectdavid-core-inference-worker
• thanosprime/projectdavid-core-training-api
• thanosprime/projectdavid-core-training-worker

All images are published automatically on every release of the source repository.

---

Related Repositories

Repository	Purpose
projectdavid-core	Source code for the platform runtime
projectdavid	Python SDK — start here for application development
reference-backend	Reference backend application
reference-frontend	Reference frontend application

---

Privacy

No data or telemetry leaves the stack except when you explicitly route to an external inference provider, your assistant calls web search at runtime, one of your tools calls an external API, or you load an image from an external URL.

Your instance is unique, with unique secrets. We cannot see your conversations, data, or secrets.

---

License

Distributed under the PolyForm Noncommercial License 1.0.0. Commercial licensing available — contact licensing@projectdavid.co.uk.