I build tools for agents, systems they can operate, and products people can inspect.
The work is connected by a practical question: how much useful work can continue without making the person at the keyboard carry every handoff?
| Sheet | Subject | What to inspect |
|---|---|---|
| 02 | The working loop | Tools, delivery and human attention |
| 03 | Enterprise systems | Domain boundaries and responsibilities |
| 04 | Inspectable products | Public source and a live explorer |
Agents need useful tools, a reliable return path, and a clear way to ask for the human. These are separate jobs.
Small command-line tools make recurring engineering tasks concrete: read work items, delegate a bounded reasoning task, and inspect structural duplication.
The audited snapshot includes linear-reads, afford-spark and afford-dry. Their interfaces are the useful unit to inspect.
Read the command contracts and examples. This is the most direct entry into the workflow primitives.
Hive connects development sessions across machines and providers. A broker carries work and responses; local edges connect those deliveries to the sessions that can act on them.
Attention is a human-intervention queue on that foundation. An agent can ask for a decision while the human keeps working.
See the original interactive specification, then the current operating design.
Try the original simulatorRead the current designRead Hive source
The Weave is the team and its working agreements: named responsibilities, task dispatch, independent review, delivery and lessons carried into later work.
Hive handles coordination. The Weave supplies the engineering practice around it.
Follow how a change moves from an assigned task to review and delivery.
The agent continues. The human chooses when to switch again.
The 15 October notification amendment adds one Slack direct message when a request is created, with later updates in place. It preserves explicit switching and does not take focus.
The same engineering questions become sharper when agents touch company knowledge, accounting, inventory and business rules.
An enterprise AI department: connect company systems, build a typed operational model, evaluate business rules and carry out governed workflows.
Source ingestion, ontology, reasoning, authorization and execution have distinct responsibilities.
How company information becomes an operational model and an authorized action.
Open the Sókrates project sheetA family of operational kernels for accounting, workforce, commitments, payroll, payments, inventory and work records.
Each domain keeps a narrow core. Shared contracts cover persistence, composition and conformance; customer policy can live above those cores.
What an agent can preview, retry and trace when a business operation has consequences.
Open the Krepis project sheetA typed interface grammar with renderers and tooling. Interfaces are represented as data that a host can validate and render.
It supplies a shared presentation layer for business systems, including the kernel family.
The separation between a structured interface description and its rendered surface.
Explore Morphe sourceThree concrete ways into the work: a creative workflow, generation controls, and a live model you can question.
A native macOS workflow for turning photographs or video clips into a diffusion-training project. Preparation, concept-focused captioning, critique, training and model collection belong to one visible process.
Inspect how concept boundaries become captions, how the user reviews them, and how local preparation connects to rented compute.
Available as a source development preview.
Experimental ComfyUI extensions for MiniMax H3 audiovisual generation: attention controls, synchronized audio/video context windows, weighted prompt timelines and shared multimodal references.
The repository documents experimental nodes and their operating requirements.
An Icelandic apartment-valuation explorer. Choose a property and valuation date, inspect the price breakdown, then change a physical fact and compare the resulting estimate.
A common market starting value is adjusted by three fitted layers. The interface retains the distinction between registered facts, listing claims and a scenario you change.
Scenario differences describe the model’s response; they do not establish renovation returns.
This map draws on the repository inventory captured on 13 October 2026, with activity counted from 13 April 2025. The linked product surfaces and notification note reflect subsequent work identified above.
Source counts in the original inventory excluded dependencies, identified generated output and mirrored plugin trees. A second total removed byte-identical source files across repositories. Test counts described declarations, not passing executions; commits and merged changes described activity within the stated window.
Repository descriptions came from source trees, manifests, readme files and implementation-bearing branches. These definitions keep the snapshot comparable without making code volume the measure of a product.