Skip to content
Portfolio / Local AI lab
Independent lab / Applied AI · running daily since July 2026

Curiosity, running
on my own hardware.

Nyx OS is my local AI partner. An open-weight model is served from a 128 GB DGX Spark on my desk; an agent answers on Telegram and in a console; and a growing set of tools lets her read a Power BI model, draft a chart, review a document or prove a change before it ships. Nothing she reads leaves the machine.

Watch Nyx work · 56 seconds ↓See a measured trial · 90 seconds ↓Discuss a workflow →
NYX / 56-SECOND SHOWCASE

From raw
to ready.

56 seconds · narrated, captions on, sound starts muted · the report at the end is a real saved file, built from synthetic data. Screen interactions are staged for the camera; the one-request flow the film illustrates is still being qualified.

Read the film outline

Bring a file and start with the question that matters. Check the costs, flag the cash risk, and make the result ready to review. Keep the facts; give them a form you can use. Three report styles from the same figures: an executive brief, a studio page and a control sheet.

THE QUESTION

Copying information between an AI chat and working tools leaves the person doing most of the task and checking every connection. Company material also cannot go to a cloud assistant.

WHAT RUNS TODAY

A local model with native vision, voice in and out, an agent on Telegram and in a console, read access to Power BI models, chart and dashboard drafting, a document reading desk, image generation, and proof tools that rehearse a change on synthetic rows before anything live is touched.

WHAT YOU CAN USE

A pattern for giving an assistant real tools without giving it your data: every tool ships with a drill that must be able to fail, and the recorded run is the evidence.

How it works · for a closer look

Nyx runs on hardware I own, behind a private network. The model answers through an open-source agent framework that holds the tools; each tool is a small, reviewed program with its own tests and a drill the agent must pass on a synthetic case. Capabilities listed on this page are live and have a recorded proof run. Not yet live: mail, spreadsheets and a desktop interface. Public films use synthetic data only.

01 / INSIDE THE LAB

Room to experiment.
A reason to connect.

The brain

One desk-side machine.

  • 128 GB DGX Spark, serving an open-weight model with native vision as the daily driver
  • Speech in and out; a wake word on the desk
  • Image and video generation on the same machine for design work
The hands

Tools with a job to do.

  • Reads Power BI models on disk: which source feeds which page, measure by measure
  • Drafts charts and dashboards from a brief in the house style
  • Reviews a document locally and shows the exact passages behind each answer
The proof

Rehearse before you ship.

  • A splice proof stages a reporting change on synthetic rows, compiles it, checks row-level security and renders the page, without writing the live model
  • Every tool has a drill planted with a known fault; a drill that cannot fail is not evidence
  • Runs are recorded with timings, so a claim can be checked against its tape
NYX / A MEASURED TRIAL · 90 SECONDS

From request
to report.

Recorded 4 October 2026 · seven fictional source files prepared and reviewed by a second agent · Nyx built the people dashboard in 77 seconds with one tool call, and 128 native checks passed. The first blocked attempt took nearly six minutes; a guided retry stopped in 14 seconds. Both failures stay in the record. Desktop acceptance of the packaged Power BI project remains unverified; nothing was published.

Why keep the failures in

A film that only shows the successful run would overstate the tool. The measured record keeps the blocked attempt, the retry and the final build side by side, so the 77 seconds can be read against what it took to get there.

02 / FROM QUESTION TO WORKFLOW

An idea becomes useful
when the pieces work together.

Every capability follows the same route from an idea to something I trust.

01 / FRAME

Start with a question.

Decide what needs to be understood, which inputs matter and what a useful result would look like.

02 / CONNECT

Give her the tool, not the data.

Write a small reviewed program the agent can call, with its own tests, and keep private material on the machine.

03 / PROVE

Make the run inspectable.

Plant a fault, run the drill, keep the recording. A tool joins the roster when its drill passes and fails for the right reasons.

Dated satellite evidence of development at Shura IslandSHURA ISLAND · 05 APR 2025 · VANTOR / ESRI WAYBACK
03 / EXPLORE A WORKING IDEA

The same curiosity.
A different perspective.

For a public example of making evidence easier to inspect, open Groundwork and compare the same place on two dates. Look for one visible change, then check what the imagery can and cannot tell you.

Explore Groundwork →
04 / KEEP EXPLORING

A question worth asking.
Something useful to build.

See that approach applied to Traza ↗Discuss a similar challengeBack to selected work