Systems

The infrastructure behind the work

Distributed compute, vector memory, and embodied AI — running on commodity hardware, entirely on-premise.

Pipeline

From event to memory

Real-world events follow a single path into queryable storage.

Capture High-speed video, document scanning, conversational audio — physical events are captured at the source
Normalize Frame extraction, OCR (Tesseract), speech-to-text (Whisper). Raw input becomes structured text
Embed BGE-large-en-v1.5 generates 1024-dimensional dense vectors. Temporal decay score (gamma_t) computed at write time
Store Milvus vector database with HNSW indexing. Cosine similarity. Gamma_t stored as scalar field alongside each vector
Retrieve Single-pass query: similarity * gamma_t. Temporally coherent ranked list. No post-retrieval re-ranking
Embodied AI

Five layers of presence

No data leaves your network. No cloud required.

Presence Intel RealSense depth cameras detect who's in the room, where they are, and when they approach
Listening faster-whisper with Silero VAD. CUDA float16 inference, sub-500ms per utterance. Emotion detection via wav2vec2
Thinking Multi-tier LLM routing: 1.5B–7B for fast response, 22B–70B for depth. Cloud fallback (Claude, Gemini) for factual queries
Speaking Orpheus 3B neural TTS. SNAC decoder, ~2s synthesis. Natural prosody and expressive range
Embodiment Unreal Engine MetaHuman. Photorealistic rendering with synchronized lip movement, facial expression, and eye contact
Infrastructure

The cluster

Distributed across commodity server hardware. Each node has a dedicated role.

Compute

Dell PowerEdge rack servers (R620, R630, R730, R840) and Supermicro nodes. Dual-socket Xeons. The pool now includes two PowerEdge C4130 GPU nodes — g806 and AG3065 (codename “Agrippa”) — each with four NVIDIA Tesla P100 16GB cards, contributing 128GB of HBM2 between them. GPUs across the cluster range from Tesla P40 (24GB) and P100 (16GB) to RTX 4080 (16GB) depending on workload.

Networking

Dell S6010-ON 40GbE switch with RoCE fabric (VLAN 100). Mellanox ConnectX adapters. Jumbo frames (MTU 9000) for inter-node vector operations.

Memory & Storage

Milvus standalone with etcd and MinIO backing. PostgreSQL for structured metadata. BGE-large embeddings (1024-dim, HNSW, cosine). TDR gamma_t on every record.

Ingestion

Artifact Ingestion Nodes (AIN) watch for incoming data — scanned documents, captured frames, conversation logs. OCR, embedding, and storage happen automatically.

Read the research behind the systems

TDR scoring, hyperdimensional computing, and the theoretical foundations.

Research