The bench

Projects

One system live in production, eight more in development. Click any project to read what it is meant to be and how it is meant to operate. Every number here is measured from the repository, not estimated — including when each one last moved.

Shipped

01Live

Deployed, publicly reachable, and still shipping.

Live in productionLast moved Aug 22· 330 commits / 30d

Nocturn

Real-Time Global Tracking & Data Fusion

A live, publicly deployed situational-awareness platform that fuses free public data feeds into a real-time interactive 3D globe, and carries provenance on every single observation.

TypeScript
React
CesiumJS
Node.js
DuckDB
Public data feeds

Lines of code: 241,850Lines of code
Commits: 928Commits
Files: 939Files
Building since: Jun 2026Building since

Read the detailsOpen Nocturn

In progress

02In Development

Ordered by where the effort is going. The production suite is the one we push hardest, and the orchestration engine is what builds it.

01 / 08Last moved Jul 26In development

Software Production Suite

One platform combining an AI-native build environment, a vetted module library, a peer-reviewed research layer, a translated collaboration space, and a bounty economy that pays in cash or shares.

Lines of code: 145,595Lines of code
Commits: 121Commits
Files: 849Files
Building since: Apr 2026Building since

Next.js
TypeScript
Turso
Drizzle
NextAuth
Vercel

Read the details

What the lab built

Nine working rooms: Workspace, Crucible, Greenhouse, Library, Tasks, Babble, QA, Telemetry, Bounties
The Crucible, a two-stage quality gate mapping code against ISO, OWASP, HIPAA, PCI DSS, FedRAMP, NIST and GDPR
Share-backed bounty contracts that vest real equity on reviewer sign-off
Passive translation: messages translate when read rather than when sent

02 / 08Last moved Sep 3· 985 commits / 30dIn development

AI Orchestration Engine

Multi-Agent Orchestration Engine

A local-first engine that routes work across local and cloud models while keeping context alive between them. The largest codebase in the lab, because it is the thing that builds everything else.

Lines of code: 679,999Lines of code
Commits: 1,154Commits
Files: 2,336Files
Building since: Mar 2026Building since

Python
TypeScript
Neo4j
Tree-sitter
Redis
Ollama

Read the details

What the lab built

Four-role dispatch model routing each task to the right class of model
Shared context store so context survives a model swap mid-task
Code knowledge graph built on Neo4j and Tree-sitter for repo-aware reasoning
Local-model execution through Ollama, so the free tier costs nothing to run

03 / 08Last moved Sep 3· 126 commits / 30dWorking prototype, in use

Personal Work Agent

Voice-First Personal Work Agent

An always-on assistant that runs on the lab's own hardware. You talk to it or type to it, from a desk, a phone, or a headset; it reads files, runs commands, and searches under explicit consent gates; and every reply is checked against a record of what actually ran before it reaches you.

Lines of code: 59,131Lines of code
Commits: 126Commits
Files: 199Files
Building since: Aug 2026Building since

Python
FastAPI
WebSockets
NATS
llama.cpp
Local models

Read the details

What the lab built

A per-turn action record that every reply is verified against before it is delivered
Consent gates with a decline default: no answer in two minutes means no
Local-first routing with measured escalation, so the private tier does the work it can prove it can do
Background dispatch that narrates from real tool events only

Working today

Voice and text conversations answered by local models, with a cloud tier only when a turn earns it
Sixteen gated tools behind consent approvals — no answer within two minutes means no
Every claim of work checked against the action record before it is spoken

Next gate

A full week replacing ad-hoc AI tabs in daily use, with every dependency on one always-on machine

04 / 08Last moved Jul 8In development

AI Interface Guide

Spatially-Aware AI Interface Guide

The assistant that popped up uninvited and understood nothing, rebuilt with the technology it always needed.

Lines of code: 11,027Lines of code
Commits: 16Commits
Files: 83Files
Building since: Jul 2026Building since

TypeScript
React
Canvas
Framer Motion

Read the details

What the lab built

Live DOM structure parsing so the guide knows what is on screen
Spatial awareness, responses anchored to the element the user is working in
Animated presence layer that stays out of the way until it is useful

05 / 08Last moved Apr 3In development

Technology Knowledge Graph

Civilizational Knowledge Monitor

An interactive knowledge graph mapping technology from foundational principles to speculative breakthroughs across 21 domains, with automated ingestion and AI classification.

Lines of code: 28,890Lines of code
Domains mapped: 21Domains mapped
Files: 112Files
Building since: Apr 2026Building since

FastAPI
PostgreSQL
Neo4j
Cytoscape.js
Ollama

Read the details

What the lab built

Automated ingestion pipeline across 21 technology domains
AI classification layer that places new work into the dependency graph
Cross-domain bridge detection, the connections nobody is looking for
Interactive graph exploration built on Cytoscape.js

06 / 08Last moved Feb 23In development

Probabilistic AI Research

Probabilistic Intelligence Research

An internal research programme storing model output as distributions with confidence scores, tracked against real outcomes. Not a product and not a revenue stream.

Lines of code: 160,560Lines of code
Files: 467Files
Status: Active researchStatus

Python
FastAPI
PostgreSQL
React

Read the details

What the lab built

Probability distributions with confidence scoring instead of point estimates
Outcome tracking that scores past predictions against what actually happened
Uncertainty propagation through multi-step reasoning chains

07 / 08Last moved Jul 27Working prototype

CAD-to-Headset Bridge

Spatial CAD Review & Markup

A real-time bridge between a running Autodesk Inventor session and an Apple Vision Pro or iPad. The model leaves the desktop and becomes something you walk around, measure exactly, and mark up with a pencil — with every change flowing back over the private network, never through anyone's cloud.

Lines of code: 65,649Lines of code
Commits: 20Commits
Files: 256Files
Building since: Jul 2026Building since

C#
.NET
TypeScript
WebSockets
Autodesk Inventor
visionOS

Read the details

What the lab built

A live bridge out of a running Inventor session, not an export pipeline
Exact-measure and markup workflows a reviewer can trust for real decisions
Local-network-only architecture: no cloud, no account, no third party

Working today

A working Inventor add-in, local hub, and viewer covering models, parts, assemblies, and drawings
Pencil-first markup and exact measurement on the headset and the iPad
Everything on the private network — the model never leaves the building

Next gate

Packaging: licensing, managed updates, and device onboarding a customer could do alone

08 / 08Last moved Aug 23· 1 commits / 30dWorking prototype

Tactical Planning Workstation

AI-Augmented Mission Planning

A local-first planning workstation for the people who still plan operations in PowerPoint over a screenshot of Google Earth. A staff of specialist planning agents — terrain, threat, courses of action, orders — do their work directly on an interactive satellite map, and produce the documents a trained planner would recognize.

Lines of code: 24,941Lines of code
Commits: 2Commits
Files: 136Files
Building since: Mar 2026Building since

TypeScript
React
Node.js
Satellite imagery
Terrain data

Read the details

What the lab built

The staff-section model: specialist planning agents in place of one general-purpose assistant
Structured planning output rendered onto the map, not into a transcript
Local-first from the start, because plans are nobody else's business

Working today

The full planning sequence on a live satellite map — area of operations through recon, terrain, threat, courses of action, and orders — each stage gated on the one before it
Recon that sweeps roads, structures, obstacles, and elevation inside a drawn boundary, with ground-level imagery probes
Nothing advances by itself: recon products stay preview-only until committed to the baseline

Next gate

One plan end to end: terrain analysis to a finished order, on one map, reviewed by someone who has written real ones

On the labelsThe orchestration engine is internal infrastructure, it powers the lab and is not sold on its own. The probabilistic AI work is internal research, not a product and not a revenue stream. Both are listed because they explain how the rest of the work gets done. There are no source links here. Code is shared inside the lab once it has been vetted, and opened publicly after that.