I build self-healing systems with as little AI as the job allows.

Twenty-seven, in Verona. Three systems — in Rust, TypeScript and Python — on a mesh of machines I own and operate. One of the three generates that mesh and runs it: plug a machine in and it configures itself. The pattern is always the same: declare the intended state, measure the real one, make the gap visible, correct it. Every layer that can be deterministic is deterministic. A model gets called only where nothing cheaper works, and it is an interface, never the foundation.

Two computers in a house in Verona serve praxisweb.xyz, hold its database, take live card payments and replicate to each other.

No cloud provider is involved. When something breaks at three in the morning, the person who fixes it is me.

Praxis — a goal journal and accountability platform. Live with 45 users, 3 on paid plans. A clinical psychologist runs it with a group of socially withdrawn young adults: shared task delegation and peer verification turn daily self-care routines into commitments someone else can see and confirm. Daily brief in the morning, one-tap check-in, recap at night. Installable, and it keeps working with no network — client state is written offline first and reconciled deterministically, so the same edit sequence converges to the same state on every device. Android shipped through 3.7.0 in 7 languages, plus SwiftUI iOS, an Electron desktop build, and an MCP server that exposes the journal to agents. Logged reality, not intention, is what revises the plan.

TypeScript · React 19 · Express 5 · Supabase/Postgres · Stripe · Offline PWA · Android · SwiftUI

physis-core — a Rust engine that reduces a corpus to structure: what repeats, what differs, what contradicts. Competing hypotheses stay side by side instead of collapsing early, and every claim carries replayable provenance back to its evidence, so a wrong answer is traceable rather than mysterious. Hybrid BM25 + cosine retrieval under a fixed token budget; embedder, n-gram table and tokenizer are swappable interfaces; runs fully offline. One engine, four surfaces — CLI, HTTP server, MCP server, TUI.

410 downloads on crates.io (721 across both crates), used autonomously by researchers outside the project: in optimization, for the predrafter and checker; in alignment, for neurosemantic orchestration and context compilation, where the model stays conceptually sandboxed and the structure around it stays inspectable.

Measured rather than asserted: semantic retrieval hit@5 0.57 vs 0.14 lexical baseline; proposal ranking top-3 0.712 vs 0.136 permuted-label null — published as triage, explicitly not certification. Negative results stay in the record. A measured failure rewrites the structure that produced it.

Rust · Apache-2.0 · ONNX · MCP

aion — the mesh. Not a controller sitting beside one: aion generates the mesh and runs it. RandoMesh is that mesh written down — documentation, not a second codebase.

Plug a machine in and it configures itself. aion provisions the control plane over SSH with idempotent installers, systemd units and timers (disable is deliberately not uninstall), then owns the lifecycle of everything running there — Praxis, physis, llama-server inference nodes, gateways. Topology is emitted into the fleet config and read back on every probe, so operator intent beats built-in defaults and drift shows up as a diff instead of a stale HUD. Zero registry by design: an instance's name derives its port, discovery is a file heartbeat, and there is nothing to register a new node with — no daemon, no mDNS, no port scan. Cross-node transport is an SSH tunnel rather than a hand-rolled TLS/PKI stack: encryption, host authentication and per-peer keys for no new protocol code.

One cockpit probes every service on every node — TCP, HTTP health or systemd — and drives it. One Intent type for every input device: keyboard, gamepad, voice, a custom hardware deck. Every backend a swappable harness. A dead node is a state, not an outage: the mesh reports it and routes around it.

Python · asyncio · SSH · systemd

Forty-five public repositories since 2022, in TypeScript, Rust, Kotlin and Python.

CAD designer for industrial food-processing machinery, then two years selling custom 3D-printed parts in engineering-grade materials — CAD to material choice to finishing to delivery. ITS Meccatronico: PLC logic, device control, low-level C on 8-bit PIC microcontrollers. Philosophy at Verona, in progress — symbolic logic, semantics, ontology. Hardware is never the ideal on paper, and neither is a deadline.

I take contract work alongside Praxis. I am most useful where one person needs to own a system end to end — from the database to the deploy — rather than fill one seat on a large team.

• Full-stack TypeScript — React, Express, Postgres
• Self-hosting, multi-node operation, migration off platform bills
• Offline-capable PWAs and Android shells
• LLM plumbing — routing, fallback, cost control, retrieval that is measured rather than assumed

praxisweb.xyz/me · [email protected] English C2 (Cambridge CAE) · Italian native

Tell me what is broken or what you want built. I answer within a day, and I will say so plainly if it is not something I should take.