Systems that remember
PRAXIS is an independent research lab building systems that capture, encode, and retrieve real-world physical events as persistent, queryable memory.
New Jersey. Open source. On-premise.
What we're building
Active projects across physical capture, memory architecture, and embodied AI.
Physical Event Capture
High-speed video capture of formation events — glass, laser ablation, fracture. Frame normalization, synthetic stereo, and structured encoding into queryable memory records.
Vector Memory with Temporal Decay
Native TDR scoring for Milvus. Significance-weighted decay computed at write time, applied at retrieval. No post-retrieval hacks. One pass, temporally coherent results.
Distributed On-Premise Infrastructure
Multi-node compute cluster running embedding, inference, speech-to-text, text-to-speech, and MetaHuman rendering across commodity server hardware. 40Gb internal fabric.
Embodied Conversational AI
Photorealistic MetaHuman with real-time speech, listening, and persistent memory. Runs entirely on local hardware. No cloud, no subscription, no API dependency.
Intelligence, embodied
The research feeds a working system — an on-premise AI with a face, a voice, and a memory.
Presence
Depth cameras detect who's in the room, where they are, when they approach. No buttons, no commands.
Listening
On-device Whisper speech recognition. Speak naturally, from anywhere in the room. Sub-500ms latency.
Thinking
Local large language models, 7B to 70B parameters. Multi-tier routing. Conversations build persistent context.
Speaking
Orpheus neural TTS. Expressive prosody, natural pacing. Not a recording, not a robot.
Embodiment
Unreal Engine MetaHuman. Photorealistic face with synchronized lip movement, expression, and eye contact.
Private
Everything runs inside your network. Your conversations, your memories, your data — none of it ever leaves.
tdr-vector-decay
Native temporal decay scoring for vector database retrieval. The TDR equation, Milvus integration patterns, and working examples — released under Apache 2.0.
Current work includes processing high-speed capture of physical formation events — including glass, laser ablation, and fracture — into structured memory records.
Latest dispatch
2026-04-30 · Infrastructure
AG3065 joins the cluster
A second Dell PowerEdge C4130 — codename “Agrippa” — came online this week alongside g806. Four NVIDIA Tesla P100 16GB cards add another 64GB of HBM2 to the compute pool, earmarked for LoRA fine-tuning and dedicated persona inference. Welcome aboard.
We explore.
Get in touch
Research inquiries, collaboration proposals, and technical discussion welcome.