Praecise
Menu
Praecise(01) An AI lab(02) A cloud

An AI lab, and the cloud to run it on.

We publish open-source systems for building and running AI: a framework for agents and an acceleration layer for inference. We operate the cloud that deploys them: apps, agents, models, data and machines, on hardware we own.

The LabThe Cloud

01

Two halves, one app

What runs in the open runs the same on the Cloud. The config gets shorter; the code does not change.

Lab · Open source · Apache-2.0

Everything measured, negative results included.

Praecise Harness makes an agent work: models, escalation, grounding, memory, workflows that stop for a person and resume. Praecise Engine makes a model faster: whether to speculate, which drafter, how to shape a batch, whom to admit. Neither needs the other. Both are on GitHub.

See the Lab

Cloud · Platform · Hardware we own

One key provisions it. One command places it.

Apps, agents, models, data, machines and automations. An app from the Harness runs on the Cloud with one key in its environment. A folder, a container or a static export is built, placed behind our edge and supervised with one command.

See the CloudOpen the Console

02

Systems(3)

Two are open source and complete on their own. The third is the managed side of the same app.
Open source · Framework

Praecise Harness

The framework for AI agents. A folder is an app. Write what an agent is for, what it knows and what it may act through. Which model runs, when to escalate, how to ground an answer, how to remember, how to resume a workflow that stopped for a person — derived, not configured.

FrameworkProtocolsDataAgents
Open source · Inference

Praecise Engine

A backend-agnostic acceleration layer for large language models. Sits between an application and an inference backend and makes token generation faster: speculation policy, drafter selection, request shaping, admission. llama.cpp is the first backend; served runtimes are reached over OpenAI-compatible HTTP.

InferenceMeasurementModelsMachines
Platform

Praecise Cloud

The managed side of the same app. One key, and the models behind the ladder, the database and object storage are provisioned and reachable through the code you already wrote. One command, and an app, an agent or a service is built, placed and supervised on machines we operate.

DeploymentDataAppsAgentsMachines

03

What the Cloud runs(6)

Six nouns. Each is provisioned through the platform and recorded by it, never placed by hand.
  1. 01

    Apps

    A folder, a container or a static export.

    Built on our machines from your source, placed behind our edge with a certificate issued, and supervised. Node, Python and static sites are recognised from the tree; nothing to describe twice.

  2. 02

    Agents

    Agents as a service.

    An agent from the Harness is served as a chat UI, a REST endpoint, an MCP server and an A2A card. Run yours, or ask us to run one for a job you name.

  3. 03

    Models

    A ladder, not a model id.

    Fast, balanced, best. The models behind each rung are ours to keep current, accelerated by Praecise Engine on our hardware. Point the same app at your own endpoints instead and nothing else changes.

  4. 04

    Data

    Postgres, vectors, time series, key-value, objects.

    A tenant is provisioned with one command and the app is handed its credentials. Vector and time-series extensions where the licence allows them; object storage for what does not belong in a row.

  5. 05

    Machines

    Virtual machines and GPU compute.

    For workloads that need a whole box: a VM with a fixed shape, or a GPU node in the class the Engine is tuned for. Hardware we own and operate, provisioned in the region a workload needs.

  6. 06

    Automations

    Workflows that wait.

    Scheduled or event-driven runs built from the Harness's durable workflows: they pause for a person, resume where they stopped, and leave a journal that says what happened.

How a deploy works

04

Research(8)

Notes from the Engine, each cited to a measurement or a primary source. The retired ideas stay published.
  1. 01

    A serving layer that admits on predicted end-to-end time, from a cost model that never saw the output length, refuses almost everything on a slow decoder, and refuses more the longer it runs. Observed on a DGX Spark at 31.6 tok/s. Fixed with an output-aware cost model and a queue-wait SLO.

  2. 02

    The first non-simulated measurement in this work. N-gram self-speculation on an RTX 5070 Ti: 556 of 598 drafted tokens accepted, 93.0% overall, and 100% at the first five positions. The acceptance curve, not the block size, is what decides the win.

  3. 03

    Same machine, same model, eight prompts. SQL 3.5×, code 1.5×, and zero drafted tokens on JSON, number lists and prose. Whether to speculate is where the value is, and the drafter makes that call correctly on its own.

  4. 04

    Decode on a DGX Spark runs at about 80% of its 273 GB/s memory bandwidth, inferred from tokens per second times bytes per token rather than read from a datasheet. The highest fraction of peak of any comparable machine, and the constraint that bounds every host-side optimisation to the remainder.

All notes

05

Start with a folder

Node 22 or newer. One dependency-free package.
$ npx praecise init my-app
$ cd my-app && npm install
$ npx praecise dev
→ http://localhost:3000/support
  chat UI · REST · MCP endpoint

A file’s name is its name: agents/support.ts is the agent support, served at /support. No registry, no imports between files, no config to keep in sync.

Put PRAECISE_API_KEY in .env and the same app runs on the Cloud. Describe your own endpoints instead and it runs on things you operate.

Read the README

06

Tell us what you’re running.

An app, an agent, a model, a dataset, or the machines you need. We are placing early workloads by conversation.

contact@praecise.com