Project NIVORA
Privacy policyCookie policy

your personalized digital lab

Project NIVORA

An artificial intelligence that works, on your machine.

Project Nivora gives a language model tools, rules and a method of work, and keeps them under the control of the person using it. CAD design, electronics, scientific computation, software development and documentary research, without a byte leaving the computer.

version 0.2.0 · alpha
Installer coming soonCode on GitHub

01What it is

Nivora is a work platform built on artificial intelligence, designed to run entirely on the user's machine. It brings language models, local or, on explicit authorisation, from outside providers, into a complete working environment for STEM activities.

Unlike a conversational assistant, Nivora produces verifiable work. Every request can become a job: a structured work order that goes through specialised benches and gives back concrete results, a CAD solid, a circuit simulation, a scientific report, a document, with the evidence of every step. Calculations and measurements do not come from the model but from dedicated engines.

Technically, Nivora is a harness: the stable infrastructure that gives the model tools, rules and a method. The model is an interchangeable component, local through Ollama or outside through an API, while the harness, with its benches, its flow of jobs, its register of activities and its security policies, stays the same.

The architecture is entirely local. A central service coordinates models and tools; the desktop panels and the local web interface connect to it. There is no remote Nivora server: no data leaves the machine except through the calls the user authorises. A disclosure page, composed from the live system, says at any moment what Nivora is made of, what leaves the machine and what stays.

02How it works

From the request to the result, with the evidence of every step.

  1. 1RequestBy voice, in the console or from a program. It becomes a job: a work order with a goal and a bench.
  2. 2GateEvery action that touches the machine is classified; what is not pre-authorised waits for a signature.
  3. 3Benches and enginesThe model coordinates; CadQuery, ngspice, KiCad, NumPy, RDKit and the others measure.
  4. 4Evidence and resultA solid, a circuit, a report or a document, with the numbers and files that prove them, in the register of activities.

03What it does

Five benches, each with its own measuring engines. The model coordinates; the engines measure.

  1. 1CADParametric solids with CadQuery, real measurements of the envelope, previews, export for 3D printing, and an external engine for the parts that ask for more.
  2. 2ElectronicsSchematics and netlists with SKiDL, simulation with ngspice, routing with KiCad and Freerouting, rule checks before a circuit is declared good.
  3. 3ScienceComputation with NumPy and SciPy, chemistry with RDKit and xtb, astronomy with Skyfield, a local corpus to search instead of remembering, reports with charts and navigable molecules.
  4. 4CodeReading, writing and testing software in separate environments, with syntax checks and tests actually run before delivering.
  5. 5ResearchDocumentary research on sources that accept programs, dossiers with citations and provenance, a browser driven and watched by the user.
The gateEvery action that touches the machine passes through a gate: read, write, network, system, destructive. What is not pre-authorised stops on an approval card that only the administrator can sign. Calls to outside providers are off by default and, when on, declared.

04What it is for, and for whom

For anyone who wants an artificial intelligence that produces measurable work and stays on their own machine.

Makers and designersA part to print, a circuit to try, a calculation not to get wrong: the request becomes a job with verifiable results.
Students and researchersAn environment where the model does not invent numbers: the engines do the arithmetic, the sources are cited, the reports can be checked.
DevelopersA harness with an interface for programs, plugins of one's own and a test suite; a workbench that can run without windows.
Workshops and practicesNivora is free at work too: MIT licence, no key, no agreement. At install time you say whether the use is personal or at work, only because two third-party components (the Marginalia index, the news feeds) are offered for personal use alone and are not installed for work.
Free for everyone, at home and at workThe code of Nivora is released under the MIT licence: use it, study it, change it and share it, in a company too, with no key and no agreement. At install time you choose whether the use is personal or at work: two third-party components, the Marginalia search index and the news feeds, are offered by their publishers for personal use alone, and for work they are not installed. The choice is shown on the disclosure page and changed by running the installer again.

05Roadmap

The milestones of the project, by version: the base for laboratories and makers, up to 1.0. Milestones, not dates.

0.2Alpha0.3Release0.4Data0.5Devices0.6Workshop0.7 – 0.9Solidity1.0Nivora

Base · to 1.0

  1. 0.2 · Alpha
    • Five benches: CAD, electronics, code, science, research.
    • Approval gate, voice and Listening.
    • Installer with the use choice and a model chosen for the machine; eight languages.
  2. 0.4 · Data backbone
    • One model for people, places, machines, items, movements and jobs.
    • Event log; backup and restore from the settings.
    • Users and roles.
    • A corpus for every bench, built by you from your own documents.
  3. 0.5 · Device bus
    • 3D printers, MQTT sensor nodes, Modbus, SCPI instruments, G-code controllers.
    • The lab as a live twin of the machines.
    • Sentinels that signal and propose; every command passes the gate.
  4. 0.6 · Workshop
    • Maintenance: plans, calibrations, spare parts tied to the inventory.
    • Inventory with locations and minimums.
    • Indicators read from the bus, not typed.
  5. 0.7 – 0.9 · Solidity
    • Updates in place, without reinstalling.
    • Full Linux; documentation for whoever installs and whoever writes a plugin.
    • Security review of what listens on the network.
  6. 1.0 · Nivora 1.0
    • A laboratory or a maker installs from this page and works connected.
    • Client and host installations complete.

06Download

The installer fetches every component from its original publisher, with pinned versions, and shows what is already on the machine and what is missing. No third-party component travels in the repository, with one declared exception.

What it needs
  • Windows 10 or 11, x64.
  • An NVIDIA or AMD graphics card: the installer reads the card and proposes the model that fits, from 2 to 30 GB of memory; with no card, on the processor.
  • About 40 GB of disk for the core, 42 with the benches.
  • A connection during the installation; afterwards, Nivora works without one.
version 0.2.0 · alpha

No release published yet: the button comes alive with the first one.

Installer coming soonCode on GitHub

07Contact

A form and nothing else: no account, no newsletter. The message reaches the people who look after the project by e-mail and is answered at the same address.