Home · Projects

Released tools and active research

What we actually build.

Explore independent projects in AI agent replay, independent verification, offline voice computing and approval-first communications. We distinguish public releases from research foundations and private experiments, with evidence and limits for each.

Our projects

Active · released Privacy & accessibility · desktop software

YazSes

An open-source, offline voice-dictation daemon for Linux, macOS and Windows. Hold a key, speak, release — the speech is transcribed on your own CPU and typed into whatever window has focus. There is no network call in the dictation path, no account and no API key, so the privacy claim is something you can verify in the source rather than something you have to trust.

It also transcribes audio and video files, and captures whole meetings as speaker-labelled transcripts with locally generated minutes. Two features exist specifically for users mainstream dictation serves badly: a Dysfluency-Friendly Mode for stuttered or dysarthric speech, and an EMG muscle-sensor trigger for hands-free use by people with motor disabilities.

Read about YazSes
Released beta AI agent evidence & replay

NovaFabric

Self-hosted, open-source infrastructure that captures AI agent executions as portable Run Capsules. It supports forensic inspection, supported mocked replay, structural comparison, provenance and optional sealing so a run can be examined after it finishes.

The responsibility question: can a team reproduce and inspect what an agent did without relying only on a hosted dashboard?

Explore NovaFabric
Research alpha Agent verification & human oversight

IDKMesh

An open-source research and engineering foundation for coordinating humans and AI agents around bounded work and independently checked results. Its current installable gate-audit command examines correlated verifier errors in AI review panels.

The responsibility question: when several reviewers approve an agent result, how independent and reliable is that approval?

Explore IDKMesh
  • Public statusResearch alpha
  • Available nowReview-gate audit CLI
  • Research targetVerified Swarm Runner
  • SourceGitHub repository
Experimental Responsible distribution & approval

ReachWeaver

Experimental, offline-first tooling and a Claude Code skill for evidence-backed communications. It prepares channel-specific drafts, keeps human approval gates and can stage scheduled work; live publishing has not been validated with connected owner accounts.

The responsibility question: how can AI help communicate real work without fabricating claims, spamming communities or silently publishing?

Explore ReachWeaver
  • Public statusExperimental research
  • SourcePrivate repository
  • Available nowLocal drafts and checks
  • Live publishingNot validated
Selection

What makes something a project here.

The lab is small and deliberately slow. Four tests decide whether an idea becomes work.

Responsibility is structural

The safety, privacy or welfare property has to come from how the system is built — something a reader can verify — not from a policy document promising good behaviour.

Someone is underserved

We prefer work that helps people or living systems that mainstream products treat as too small a market to design for.

It can be checked

Open source, published tests, benchmarks, or an archived release with a permanent identifier. If a claim cannot be independently checked, we do not make it.

A small team can finish it

We would rather ship and maintain one useful thing than announce five. Capabilities that are not built yet are published as “planned”, not described as if they worked.

In development

Further work in development.

No other work is listed as a project until its evidence and release status can be described accurately. For collaboration, get in touch.

Let's build responsible systems together.

We welcome collaboration, advisors, and partners across responsible AI, sensing, verification, and human-centered design.

Get in touch