About Noetic Synthesis
An independent research lab building sovereign embodied AI on enterprise hardware.
Noetic Synthesis was founded to answer a practical question: how do you make real-world physical events — things that happen once, in a specific place, at a specific time — queryable?
We build the capture pipelines, memory architectures, and retrieval systems that turn transient physical events into persistent, searchable records. The work spans high-speed video capture, vector database engineering, temporal decay scoring, and embodied AI.
The applied output of this research is a single MetaHuman prototype — a photorealistic, conversational AI with persistent memory, running entirely on our own hardware. No cloud. No internet required. Fully sovereign operation.
We are the Praxis Collective. Based in Cherry Hill, New Jersey. Built in America.
What we run on
This is not a simulation. The prototype runs on dedicated enterprise hardware.
10+ Enterprise Servers
Dell PowerEdge (R620, R630, R730, R840) and Supermicro (SYS-6029TP) rack servers. Dual-socket Xeons, up to 72 threads per node. Purpose-assigned: embedding, inference, STT, TTS, rendering, memory.
200GB+ Total VRAM
GPUs across the cluster: Tesla P40 (24GB), RTX 3060 (12GB), RTX 4070 (12GB), RTX 4080 (16GB), RTX 5060 Ti (16GB). Each assigned to a specific workload — no time-sharing, no contention.
40Gb RoCE Fabric
Dell S6010-ON switch with Mellanox ConnectX adapters. Jumbo frames (MTU 9000). Dedicated VLAN for inter-node vector operations and model communication.
Sovereign Operation
The entire system runs without internet connectivity. LLM inference, speech recognition, voice synthesis, memory retrieval, and MetaHuman rendering — all local. No API calls leave the network.
Active R&D
-
Time-Decayed Resonance (TDR) Native temporal decay scoring for vector retrieval. Significance-weighted, computed at write time. Open source.
-
Hyperdimensional Computing (HDC) Binary vector encoding for cross-modal retrieval. 10,000+ bit hypervectors with Hamming distance search.
-
LoRA Stacking Layered low-rank adaptation for personality, domain knowledge, and conversational style — applied to local LLMs without full fine-tuning.
-
Physical Event Capture High-speed video of formation events (glass, laser ablation, fracture). Frame normalization, synthetic stereo, structured encoding into memory.
-
Praxis Sense Integrated sensory system: depth vision, spatial audio, voice activity detection, and emotion recognition fused into a unified perception pipeline.
How we work
-
Experiment first. Theory follows from working systems, not the other way around.
-
Sovereign by default. No internet required. No cloud dependency. No data leaves the network.
-
Open foundation. Built on open-source models. Research published openly.
-
Single focus. One MetaHuman prototype. Done right.
Public work
Research we've released for the community.
tdr-vector-decay
Native temporal decay scoring for vector database retrieval. TDR equation, Milvus integration, working examples. Apache 2.0.
Praxis Collective
Organization home for all public repositories and research releases.
We explore.
Let's talk
Research inquiries, collaboration proposals, and technical discussion welcome.
Contact Us