Luca Visciola.
Full stack developer, self-taught, still a musician about it.

S-MoE — A 235B frontier model, answering from a laptop SSD. No cloud. No H100.

Seismic Mixture of Experts. A model-agnostic inference engine for fine-grained MoE models that streams experts from NVMe, so the model never has to fit in RAM.

The mountain rests cold on the SSD. Only the handful of experts each word actually needs is ever pulled into memory.

Point the shatter script at any compatible fine-grained MoE checkpoint and it produces a self-describing vault: RoPE theta, routing top-k, attention geometry and lineage all travel inside the .smoe file, so the engine reads what it is given instead of assuming what it was built for. Swap a 235B for a 30B at runtime — same binary, same laptop, ~13.7 tok/s.

Every byte, every struct field, every atomic flag is mapped, explained and public. No secrets. No priesthood. A door in the memory wall, and the map to walk through it.

Designed with an AI agent across many sessions: the direction came from human curiosity, the implementation depth from the collaboration.

ANDARTIS — A private fleet of tiny neural networks, trained only on your data.

The current AI landscape is obsessed with one model to rule them all. ANDARTIS rejects this.

Instead of a massive, slow, privacy-leaking generalist, ANDARTIS lets you build a private fleet of tiny, specialised neural networks trained exclusively on your own data. It is the evolution of ARKHEIN — away from general-purpose models, toward a distributed ecosystem of micro-models and conversational SLMs.

• Speed as a human right. Micro-models and SLMs under 8B parameters run on the Apple Neural Engine in milliseconds.
• Total data sovereignty. No APIs, no telemetry, no clouds. Your machine is the forge; your data is the fuel.
• Hybrid intelligence. The fluidity of a conversational Senior Analyst combined with the 100% precision of symbolic micro-models.

A PHP orchestrator and an Apple MLX acceleration engine talk over a zero-latency JSON-RPC STDIO pipeline. Each Intelligence Node gets its own isolated core.sqlite — Project A can never contaminate Project B. The Intent Blade picks the tool, the orchestrator fetches raw rows, the Senior Analyst turns them into prose. The numbers come from SQL, so there is nothing left for the model to hallucinate.

ARKHEIN — Command your silicon. Own your memory. A sovereign macOS agent.

Most “agents” are thin wrappers for cloud-based surveillance. They rent you your own intelligence and sell your data on the way out.

ARKHEIN is a local-first macOS agent that lives entirely on your silicon. No telemetry. No cloud gatekeepers. No data ever crosses the hardware boundary. Deep complexity hidden behind a minimalist shell.

The Vantage Hub. Knowledge is organised into sovereign silos — every authorised folder is a physical partition with its own semantic grounding, and 100% topical isolation between them.

The Mind. An eight-level cognitive pipeline — grounding, perception, context, canopy, decomposition, scratchpad, critique, generation — moves complexity out of the model weights and into the process, so small local models can follow reasoning that would otherwise need a trillion-parameter giant.

The Memory. SQLite is the single source of truth; the vector index is a disposable binary accelerator that can be rebuilt from it with zero downtime.

The Hand. Filesystem commands with surgical precision, fenced by the Silo Guard — mathematically absolute sandbox boundaries. Every action requires a verified plan. You remain the final authority.

INTENTIO — Not a chatbot. Not a cloud service. A bounded space where AI pays attention.

A local-first, private cognitive environment designed to operate inside a carefully constructed knowledge space.

It does not try to know everything. It tries to understand one context well — the one you design.

Most AI systems are built as oracles: global knowledge, generic answers, invisible data flows. INTENTIO rejects that model and treats AI instead as an interpreter of meaning inside a bounded cognitive space. Private — nothing leaves your machine. Local — models and embeddings run on your hardware. Intentional — structure matters more than scale. Inspectable — every component can be understood.

What it is not: a chatbot trained on the internet, a “chat with your docs” demo, an autonomous agent pretending to think, or a cloud-dependent service.

Private AI is not a feature. It is an architectural decision.

This is the project about attention, structure and restraint rather than spectacle — and the one that names the thread running through everything else here.

Redis Blog — Redis as the primary database, not the cache in front of it.

A modern, high-performance blog boilerplate built with Nuxt and Redis, containerised with Docker.

The premise is a small heresy: treat Redis as the primary store for a content-driven application rather than the cache you bolt in front of Postgres. Hashes for posts, sorted sets for ordering and pagination, sets for tags — the data structures do the work an ORM would otherwise do badly.

The result is a stack with almost nothing between a request and its answer. It is also the clearest example of how most of these projects start: a question about whether the default architecture is actually required, followed by a working answer.

Underwind — We love Bootstrap. We really do. So we let the wind change direction.

A WordPress starter theme built on the rock-solid _underscores foundation and inspired by Understrap.

We love Bootstrap, we really do — but we missed the freedom and joy of Tailwind CSS. So we let the wind change direction.

Underscores gives you the correct WordPress template hierarchy and none of the opinions; Tailwind gives you the opinions you actually chose. The theme is the seam between the two, kept thin enough that you can read all of it before you start editing.

HydePHP Layouts Manager — Layouts and reusable components for HydePHP, without the ceremony.

A Composer package that simplifies how layouts and reusable components are managed inside HydePHP projects — dynamic layout switching, a component registry with default config, and theme handling that stays out of the way.

Part of a small family of HydePHP work that also includes an events module, a multilanguage experiment whose README documents exactly where it broke, and Hyde Wizard, a tutorial told as a story about a doodler who becomes a best-seller.

Archived now, but this is where the habit was formed: publish the package, publish the documentation, publish the failure too.

MelaFreelance — Time tracking built by a freelancer who needed it, for freelancers who do too.

Client, project and task management with hourly rates and payment tracking — the unglamorous software that decides whether a solo practice is actually profitable.

Built for my own studio first. That is the only reason it handles the awkward parts: mixed rates on one project, work logged days late, the difference between hours delivered and hours paid.

It is also the reference for how this site ships — the same GitHub Actions workflow shape, the same preference for a build that anyone can run locally and understand.

Mela Letterhead — Write the document in Markdown. Get it back on your own paper.

A command-line tool that puts a real letterhead around a Markdown file — the band across the top of page one, the small running header on the pages after it, the footer with the legal and banking details on the last — and hands back a PDF.

The paper is described once, in a single letterhead.yaml: how many fields sit in the header, what the footer columns say, which colours and fonts it uses. Documents are plain Markdown and never contain layout, so nothing drifts between a quotation written today and one written two years ago. Every label can be written once per language, which is why the same configuration prints an Angebot and an offerta without a second copy of itself. No logo sets the brand name in type instead, at the size the logo would have had — a finished letterhead, not a fallback.

Underneath it is Pandoc and Typst, so the typesetting is real typesetting rather than HTML pretending to be a page. All resolution — language, units, defaults, validation — happens in Python, and the Typst module receives a structure with one language already chosen and every length already a number of points. It holds no brand, no language and no document, which is why it never needs editing to print somebody else’s paper.

It runs on your machine, against files you own, with no account and nothing uploaded. Invoices and contracts are exactly the documents that should never have needed a service in the middle.