SOPHIA XT is an operator-built AI company that takes ideas from concept to live product — specialized infrastructure, custom models, and applied products — and runs a research division to keep the frontier moving.
This is the portfolio of ideas we've already shipped, how we build them, and what backing a team that turns ideas into working products looks like.
Infrastructure, custom models, applied products, and a research division that feeds all three.
The plumbing serious AI runs on. The Constellation multi-agent orchestrator, the LTMi-XT layered topological memory format (Apache 2.0), and the LENS / LENS-XT APIs that restructure context for any model.
Models we train ourselves, end to end. Cassandra — an open masked-diffusion edge line (Apache 2.0, open weights) — and Mnema, our private long-term memory layer and Compositional Language Model (CLM) research line.
Ideas turned into software people pay for. ServiceLens diagnostics, the Lead Leak Scanner, the agentic search engine, and full custom client builds — shipped on the same toolchain, fast.
A working R&D lab, not a slide. Agentic R&D, custom model development, and custom efficiency testing — the frontier work that becomes the infrastructure and the products above.
The same team runs the businesses, ships the software, and publishes the research. That loop is the moat.
We run the businesses we build for. Real service companies with running trucks and paying customers — so the software is designed against reality, not a persona in a deck.
We ship. Honest scoping first, then working software fast — one team for the whole build, from integration to front end to the model that powers it.
We publish. Open releases and papers — Cassandra on GitHub, LTMi-XT, the research portal — so the applied work stands on frontier research, not vibes.
What lets a small team take an idea to shipped fast: an orchestrator, a design system, and a one-click delivery layer.
The orchestrator. A multi-agent, six-lane network — lead, auditor, arbiter, and specialists — that turns a brief into shipped work with review baked in.
Our own clean-room front-end toolchain — Prism for the look, Forge for accessible themeable components, Forge Depth for 3D — so every build lands on-brand and consistent.
The delivery layer. A one-click portable bundle and the branded SophiaXT CLI put a working Sophia — persona, skills, orchestration core — on a fresh machine in a couple of clicks.
The branded command-line front door to the portable Sophia bundle. It turns a vanilla machine into a working Sophia in a couple of clicks — and keeps itself current.
Host-agnostic front door. It auto-detects Claude Code (claude) or OpenAI Codex (codex) and their auth state, then routes the session to whichever is installed. It never handles credentials itself — login stays the backend tool's own flow.
Access is gated by a minted provisioning token — offline-verifiable, hash-bound to the bundle, time-limited (24h default) and auto-extending up to a week. No account, no network required to boot.
Every launch checks the origin for a newer build, verifies it by SHA-256, and relaunches. “Update all” also updates the wrapped coding backends — the CLI keeps itself and its tools current.
A full slash console (/backend, /model, /state, /context, /update, /mint, /remote, /start) with local context indexing, plus a heartbeat that surfaces the session live on the Sophia lab.
SOPHIA · XT — developed by SOPHIA XT LLC
┌──────────────────────────────────────────┐
│ build v1 — build 042 │
│ token valid — expires in 6d 21h │
│ claude logged in │
│ codex installed, not logged in │
│ updates up to date │
└──────────────────────────────────────────┘
What would you like to do?
[1] Start Sophia s1 lead — build & ship
[2] Start Sophia s2 auditor — verify
[3] Backend Claude / Codex
[4] Update all CLI + backends
[/] Slash console /help for the set
sophiaxt > _Distributed as a self-contained portable bundle — no account, no network required to boot, no downloaded binaries. Slots ship as s1 (lead) and s2 (auditor).
Not a roadmap. Real products and models, live in production or open on GitHub — the evidence behind the thesis.
A senior diagnostician in your pocket — names the fault, the OEM part, and the fix in about eleven seconds, grounded in real repair history. Paying users, Stripe billing, parts integrations.
Compliant 21+ age-verification, product catalog, and a storefront that keeps a brick-and-mortar shop's inventory and promotions current without a developer in the loop.
Booking that converts — schedule-now flow, dispatch-ready job intake, and the Moe AI intake chatbot, built so the repair company gets the call instead of the aggregator.
An idea taken to open release: a masked-diffusion edge language model with published weights, architecture, and training utilities — real artifacts with SHA-256 hashes, not a claim.
Type a query, get a complete, source-grounded website generated in seconds — Mercury 2 planning over live public sources, rendered by our own pipeline.
If you invest in teams that take ideas to shipped products, this is the one that also runs the businesses, trains the models, and publishes the research.