A pluricellular organism of local AIs, built to grow when hardware arrives.
Below it: MESH_SILOS, AUTO_TARDEO and four private works, all local-first.
For teams that own their hardware and their data.
Now
Pluricellular runtime · MIT · v3.4.0
Several local AIs that behave as one organism: cells that serve on their own, a tissue that coordinates them, and a growth rule that changes the shape when hardware arrives.
Rust and Tokio in the tissue, Python and PyTorch in the data plane, one Protobuf contract. The scheduler builds topologies instead of choosing between models, and nodes hand over adapters, not data.
Six projects, one shared constraint: each runs where the data already lives. The hardware is the deployment target, so there is no round trip to someone else's server and no per-token billing.
Where a number appears on this page it comes from a run a reader can repeat. Where the code is private, the card says so, and it ends up shorter than the ones that can show their work.
Releases are tarballs cut at the last commit: tests, notes and source, unpacked on a machine with no account on it.
Six projects in three tiers: the flagship above, two systems that ship, and four works that are private and say so on the card.
Supporting · data governance
Map reporting silos into one crosswalk. No data moved, no network.
Supporting · social planning
A night out that plans itself: the route stays hidden until the alarm, the bill is split by rule.
Private · edge analytics
Edge retail analytics: nodes re-identify people locally and share embeddings, not images.
Private · algorithmic trading
A colony of trading strains competing for capital, mutating when they underperform and pruned when they fail.
Live · Reading
RSVP speed reader: one word at a time, at the character the eye rests on, with AI summaries. Live at neuralbricks.org.
Private · port
PicoClaw, Sipeed's ultra-light Go AI agent, ported to Radxa single-board computers.
The flagship is documented before it is described: fifteen technical notes, a verification method and a benchmark methodology, each readable as a page and fetchable as plain text.
architecture.md What lives in Rust, what lives in
Python, and the contract between them.growth-protocol.md The score that decides when a
swarm should widen and when it should not.benchmarks/methodology.md What was measured,
on what hardware, and what the numbers do not say.
I am a senior business analyst. My work sits in the seam between finance, the business and data engineering, and has done for years across asset management, banking, aviation and real estate: regulated industries where a number that cannot be traced is a number nobody will use.
Most of the job is the unglamorous half. Requirements that survive contact with the people who have to use them. A functional specification that actually ships. The dashboard that gets used instead of the one that wins the design review. An operating model that still holds a year later, when the analyst who wrote it has moved on.
I use Databricks, AWS, BigQuery, Power BI, SQL and Python myself, because a specification nobody can build is a wish with a deadline. Fluent English, native Spanish, intermediate Portuguese.
All of it is local-first for the same reason the day job keeps coming back to: the data belongs to whoever owns the hardware. That is why MESH_SILOS refuses to open a socket, why AUTO_TARDEO splits a bill with a rule instead of a conversation, and why HECATE hands over adapters rather than the data they were trained on. When a measurement contradicts the story, the measurement stays and the story is rewritten. That is what happened to this project's own identity.