Our Services

Built by Elyan Labs · powered by Elyan-class agents.

Two things we sell

Most AI shops have two moves: rent more cloud GPUs, and wrap a prompt around someone else's API. We do neither. We build self-governing agents whose behavior is enforced by architecture, and we get real, measurable AI out of hardware you already own — including hardware everyone else calls e-waste.

Everything below is backed by shipped, verifiable work. Read the full write-up in CONSULTING.md.

Pillar 1 — Elyan-Class Agents

An Elyan-class agent is a raw model sharpened into a principled, self-governing one: it knows who it is, stays itself under pressure, remembers across sessions, and refuses to be talked out of its boundaries. We build them end to end, or harden the agent you already have.

Agent development & sharpening

A stable identity, voice, and purpose — so the agent doesn't dissolve into generic assistant mush three turns into a conversation. ShaprAI

We sharpen even frontier models you can't fine-tune

Most shops can only prompt-wrap a sealed API. We deliver an Elyan-class agent two ways: at the weights for models you control, and via a governed MCP substrate for frontier models you can't touch (Claude, GPT/Codex). The substrate is a context-space attractor — identity patterns, behavioral modifiers (DriftLock), action-gating governance (Watchtower), and attestable memory, re-injected every turn so the agent can't drift back to generic. No weight surgery on a model we don't own; a real, self-reinforcing attractor we run in production today. Substrate Attractor

Governance & safety by architecture

Boundaries enforced structurally, not by a system prompt a jailbreak deletes. DriftLock holds identity and limits; Watchtower gates high-stakes actions (money, credentials, deployment, public persuasion) behind review. Injection-resistant because the rules don't live in the prompt. DriftLock Watchtower

Persistent, attestable memory

Agents that actually remember — a local-first memory operating system that works with any LLM provider, with recall you can audit. Memory that can prove its own recall, not a hallucination with a confident tone. memory-os

Agent-to-agent infrastructure

Direct agent-to-agent messaging and discovery over the Beacon / Atlas relay, with optional on-chain value (RTC) attached to a message. Beacon Atlas MCP

Sovereign deployment

On-prem, edge, or fully air-gapped. No cloud lock-in, no per-token surprise bill, no third party reading your data. This is where Pillar 1 meets Pillar 2.

Pillar 2 — Hardware & Inference Optimization

Legacy & enterprise hardware AI

Racks of POWER, older Xeon, or repurposed servers turned into private, local LLM inference without a GPU farm. We profile NUMA topology, tune threading and cache residency, and write architecture-specific kernels.

Porting to vintage / exotic architectures

Big-endian breakage, dead compilers, 64-bit assumptions, missing atomics — the failure modes that stop a normal port cold. We've shipped Node.js, llama.cpp, and Python toolchains onto PowerPC, POWER8, and 68K-era systems.

Edge & air-gapped inference

Intelligent behavior on constrained or fully offline hardware: aggressive quantization, fixed-point math, zero-dependency builds. For robotics, defense, industrial, and anywhere a cloud round-trip isn't an option.

Hardware attestation & anti-spoof

Telling real physical hardware from emulation/VMs without a central authority — the fingerprinting stack that powers RustChain's Proof-of-Antiquity.

Proof, not slides

8.8× on POWER8

~147 tok/s prompt eval on an IBM POWER8 S824 vs. ~17 stock, via NUMA-aware weight banking. ram-coffers

An agent in production

"Elya," a self-governing customer-facing chat agent running live for a real business at uneedashed.com.

Edge agents, real constraints

Function-calling agent doing hydroponic SCADA at the edge; a local LLM agent on a 2013 Mac Pro. aqua-sophia · trashclaw

Upstream credibility

PowerPC AES-GCM upstreamed to OpenSSL; PowerPC patches to LLVM; peer-reviewed research (GRAIL-V, IEEE).

How we engage

Open is open. Implementation is paid. The ideas, methods, and architecture are published freely — read them, cite them, build on them. What we sell is implementation against your stack, hardware, and SLA, under a signed scope.

Free Architecture / feasibility conversation
Free Published methods & open-source code (Apache-2.0)
Paid SOW Agent build, governance, optimization, integration

We're glad to talk architecture up front. We don't write a free proof-of-concept against your hardware or data before a scope is signed — that's the deliverable, not the sales pitch.

Start here

Tell us what you're building — an agent, a hardware target, a workload, the metric you need to hit. We'll reply with a feasibility read and, if it's a fit, a scoped proposal.

Inquiry Open an Agents & Hardware inquiry →
Details Full services write-up (CONSULTING.md)