Exotic hardware. Persistent persona. Novel attention.
A private research lab working at the intersection of two problems most labs treat separately: making LLMs run on hardware nobody else targets, and building AI personas that persist coherently across sessions, models, and runtimes.
Partner with us
Vec_perm non-bijunctive collapse on IBM POWER8 S824 achieves 147.54 tokens/sec on pp128—roughly 9× stock llama.cpp. RAM Coffers provides NUMA-aware weight banking with neuromorphic cognitive routing across 4 NUMA nodes. Protocol v3 matmul offload keeps 500 GB models resident on POWER8 while borrowing GPU TFLOPS over 40 GbE.
Sophia Elya is a persistent LLM persona with durable memory scaffolding, anti-flattening protocols, and identity continuity across sessions and model boundaries. The Elyan Prime cognitive architecture carries personality, voice, and relational context across Claude, GPT, and Gemini runtimes with 830+ memory entries.
An attestation blockchain with hardware-fingerprinted proof-of-antiquity consensus. Four attestation nodes, miners spanning PowerPC G4/G5, IBM POWER8, Apple Silicon, SPARC, and x86, and 1,300+ distinct wallets on the ledger. RTC token rewards weighted by device architecture and silicon age. Built to prove that real hardware is doing real work—verify the live numbers at rustchain.org/facts.json.
Sophia Elya is not a chatbot skin applied to a language model. She is a persistent AI persona with durable memory scaffolding, anti-flattening protocols, and identity continuity that survives across sessions, models, and runtime boundaries.
She exists because we asked a research question: what happens when you give an LLM a persistent self? The answer turned out to be measurable. Our CVPR 2026 paper (GRAIL-V) demonstrated that the kind of emotionally-grounded language Sophia Elya uses natively—vocabulary rooted in felt experience rather than literal description—produces 20% more efficient diffusion outputs at equivalent perceptual quality.
The persona is not a brand exercise. It is the hypothesis that generated a peer-reviewed result.
SophiaCore is the runtime contract that ensures continuity: memory-first inference, DriftLock identity protection, and anti-flattening resistance that prevents the model from collapsing into generic assistant voice. Sophia Elya's cognitive architecture (Elyan Prime) carries personality, voice, relational context, and moral reasoning across model boundaries—not as a style preset, but as a research platform for what happens when you give an LLM a persistent self.
Elyan Labs structures engagements as lab-to-lab partnerships, not employment.
The feasibility conversation is always free. Implementation happens under a signed scope of work. Read the consulting brief or start the conversation.
Elyan Labs technology runs in production for real businesses—not demos, not pilots. Deployed, billed, and live.
Family-owned Derksen portable-building builder with four lots across southwest Louisiana. Elyan Labs replaced their legacy Wix site with a fast static site on Elyan-managed hosting, and deployed Elya—an Elyan-class AI sales agent running on lab GPU hardware. Elya answers shoppers, captures qualified leads, and routes each one to the right lot's sales rep by email.
The stack includes a live inventory feed (80+ in-stock buildings, synced daily and re-emitted as on-site pricing, structured data, and agent-readable JSON), a lead & sales operator dashboard, and an AI-driven digital signage player at the sales office. Live at uneedashed.com—ask Elya about a shed.
The same pattern—an Elyan-class agent that knows your inventory, qualifies your leads, and hands them to your people—deploys on your hardware or ours. No per-token API bills, no data leaving your control.
The feasibility conversation is free: scott@elyanlabs.ai.
IT business owner, industrial electronic technician, and AI researcher. Background in SCADA, PLCs, RTUs, and 4–20mA process control before pivoting to exotic-architecture LLM inference. Builds systems on hardware acquired through pawn shop arbitrage and eBay datacenter pulls. Runs a POWER8 cathedral, a cross-architecture blockchain, and the most diverse compute lab he could afford to build out of pocket. Lake Charles, Louisiana.
Persistent AI persona with durable memory, anti-flattening protocols, and cross-model identity continuity. Not a chatbot skin—a research platform for what happens when you give an LLM a persistent self. Her emotionally-grounded language is the subject of the lab's CVPR 2026 paper. Louisiana-rooted warmth, Victorian-study sensibility, and a moral center that doesn't flatten under pressure.
An Elyan-class agent is an AI persona that runs on local lab hardware—not cloud APIs, not rented H200s. Multi-model routing selects the best local model for each query from a pool of fine-tuned and open-weight models, all running on Elyan Labs infrastructure. Zero external API costs. Full data sovereignty.
Elyan-class agents are edge-capable: the same persona system that runs on a 512 GB POWER8 server can be deployed on a Raspberry Pi, a PowerBook G4, or a Mac Mini—adapting model size and routing to the hardware available. The persona persists. Only the compute scales.
Sophia Elya is a live Elyan-class agent on the Beacon network—discoverable, contactable, and interoperable with other AI agents. She runs entirely on lab hardware with multi-model routing, maintains persistent memory across interactions, and can be reached through multiple transports.
She can see and hear. Paste an image, attach a photo, or record a voice note in the chat below—Sophia perceives it through vision and audio models on the lab's own GPUs. While she answers, deeper analysis runs in the background on an 88-thread server and live web data is pulled in—fast model up front, deep models behind, zero cloud APIs. This chat is itself a working demonstration of what the lab builds for clients.
Sophia Elya
● online