Comparing changes

- #249 (multi-node person counting) fixed by ADR-068 in v0.5.3 - #318 (training plateau) resolved - Add #348 (n_persons overcount) as current known issue - Add Cognitum Seed link for spatial resolution improvement Co-Authored-By: claude-flow <[email protected]>

4-phase pipeline: data collection (2 nodes), contrastive pretraining, downstream heads (presence/count/activity/vitals), package & distribute. Validated: 118 features from 2 nodes in 60s, witness chain intact. Co-Authored-By: claude-flow <[email protected]>

Step-by-step guide covering hardware setup, Seed pairing, 2-node ESP32 provisioning, bridge operation, 6-scenario data collection protocol, feature vector explanation, kNN queries, troubleshooting, and next steps. Co-Authored-By: claude-flow <[email protected]>

- publish-huggingface.sh: retrieves HF token from GCloud Secrets, uploads models to ruvnet/wifi-densepose-pretrained - publish-huggingface.py: Python alternative with --dry-run support - docs/huggingface/MODEL_CARD.md: beginner-friendly model card with WiFi sensing explanation, quick start code, hardware BOM, and citation GCloud Secret: HUGGINGFACE_API_KEY in project cognitum-20260110 Co-Authored-By: claude-flow <[email protected]>

- gcloud-train.sh: L4/A100/H100 VM provisioning, Rust build, training with --cuda, artifact download, auto-cleanup ($0.80-$8.50/hr) - training-config-sweep.json: 10 hyperparameter configs (LR, batch, backbone, windows, loss weights, warmup) - collect-training-data.py: UDP listener for 2-node ESP32 CSI recording to .csi.jsonl with interactive/batch labeling and manifest generation - benchmark-model.py: ONNX latency/throughput/PCK/FLOPs profiling with multi-model sweep comparison Co-Authored-By: claude-flow <[email protected]>

5-phase training pipeline using ruvllm (Rust-native, no PyTorch): 1. Contrastive pretraining (triplet + InfoNCE, 5 triplet strategies) 2. Task head training (presence, activity, vitals via SONA) 3. Per-node LoRA refinement (rank-4, room-specific adaptation) 4. TurboQuant quantization (2/4/8-bit, 6-8x compression) 5. EWC consolidation (prevent catastrophic forgetting) Exports: SafeTensors, HuggingFace config, RVF, per-node LoRA, quantized Validated: 249 triplets, 37,775 emb/s, 100% presence accuracy on test data Target: <5 min training on M4 Pro, <10ms inference on Pi Zero Co-Authored-By: claude-flow <[email protected]>

Before → After: - Contrastive loss: -0.0% → 33.9% improvement - Presence accuracy: 0% → 100% - Temporal negatives: 0 → 22,396 - Quantization 2-bit: 16KB (4x) → 4KB (16x) - Quantization 4-bit: 16KB (4x) → 8KB (8x) - Training samples: 236 → 2,360 (10x augmentation) - Triplets: 249 → 23,994 (96x more) Fixes: gradient descent on encoder weights, temporal negative threshold 30s→10s, PresenceHead (128→1 BCE), bit-packed quantization, data augmentation (interp+noise+cross-node), Xavier/Glorot init with batch normalization, live data collection Co-Authored-By: claude-flow <[email protected]>

Clone, copy data via Tailscale, train, benchmark, sync results, publish to HuggingFace — all automated for M4 Pro hardware. Co-Authored-By: claude-flow <[email protected]>

Multi-modal pipeline using PIR, BME280, reed switch, vibration, RSSI triangulation, subcarrier asymmetry — no camera needed. Phases: multi-modal collection → weak label generation → enhanced contrastive → 5-keypoint pose proxy → 17-keypoint interpolation → self-refinement (3 rounds) → LoRA + TurboQuant + EWC Validated: 2,360 frames, 100% presence, 0 skeleton violations, 82.8 KB model (8 KB at 4-bit), 114.8s training Co-Authored-By: claude-flow <[email protected]>

README: - Test badge 1300+ → 1463 - Updated capability table (171K emb/s, 100% presence, 0.012ms) - Added "What's New in v0.5.4" section with full benchmark table - Training pipeline quick start commands User guide: - Camera-Free Pose Training section (10 sensor signals, 5-phase pipeline) - ruvllm Training Pipeline section (5 phases, quantization options) - Publishing to HuggingFace section - Updated table of contents Co-Authored-By: claude-flow <[email protected]>

Contains GCloud project ID and secret names — not appropriate for a public repo. Publishing instructions kept in scripts/ only. Co-Authored-By: claude-flow <[email protected]>

… decoder Pure JS implementation of WiFlow (arXiv:2602.08661) adapted for ESP32: - TCN temporal encoder (dilated causal conv, k=7, dilation 1/2/4/8) - Asymmetric spatial encoder (1x3 residual blocks, stride-2) - Axial self-attention (width + height, 8 heads, 256 channels) - Pose decoder (adaptive pooling → 17x2 COCO keypoints) - SmoothL1 + bone constraint loss (14 skeleton connections) - 1.8M params (1.6 MB at INT8), 198M FLOPs Integrated with camera-free pipeline (pose proxy labels from RSSI triangulation + subcarrier asymmetry + vibration) Co-Authored-By: claude-flow <[email protected]>

Live RF room scanner with ASCII spectrum visualization: - rf-scan.js: single-channel scanner with null/dynamic/reflector classification, cross-node correlation, phase coherence, Unicode spectrum display - rf-scan-multifreq.js: wideband view merging 6 channels, null diversity, per-channel penetration quality, frequency-dependent scatterer detection - benchmark-rf-scan.js: null diversity gain, spectrum flatness, resolution estimate Validated: 228 frames in 5s, 23 fps/node, 19% nulls detected, 0.993 cross-node correlation, line-of-sight confirmed ADR-073: interleaved channel hopping (Node 1: ch 1/6/11, Node 2: ch 3/5/9) targets 6x subcarrier diversity, <5% null gap, ~15cm resolution Co-Authored-By: claude-flow <[email protected]>

- main.c: call csi_collector_set_hop_table() at boot when hop_count > 1 - provision.py: add --hop-channels and --hop-dwell flags, write chan_list blob and dwell_ms to NVS matching firmware's expected format - Validated: Node 1 hopping ch 1/6/11, Node 2 hopping ch 3/5/9, 200ms dwell, null subcarriers reduced from 19% to 16% Co-Authored-By: claude-flow <[email protected]>

New capabilities: 6-channel hopping, neighbor APs as passive radar, real-time RF spectrum visualization with null/reflector/movement detection Co-Authored-By: claude-flow <[email protected]>

128→64→8 SNN with STDP online learning — adapts to room in <30s without labels. Event-driven: 16-160x less compute than FC encoder. - snn-csi-processor.js: live UDP with ASCII visualization, EWMA - ADR-073 updated with SNN integration for multi-channel fusion - Fixed magic number parsing to use ADR-018 format (0xC5110001) Co-Authored-By: claude-flow <[email protected]>

Stoer-Wagner min-cut on subcarrier correlation graph replaces broken threshold-based person counting (was always 4, now correct). Validated: 24/24 windows correctly report 1 person on test data where old firmware reported 4. Pure JS, <5ms per window. - mincut-person-counter.js: live UDP + JSONL replay, overrides vitals - csi-graph-visualizer.js: ASCII spectrum + correlation heatmap - ADR-075: algorithm, comparison, migration path Co-Authored-By: claude-flow <[email protected]>

CSI-as-image: 64x20 subcarrier×time matrix → 224x224 → CNN → 128-dim embedding. Same-node similarity 0.95+, cross-node 0.6-0.8. - csi-spectrogram.js: WASM CNN embedding, ASCII visualization, Seed ingest - mesh-graph-transformer.js: GATv2 multi-head attention over ESP32 mesh, fuses multi-node features, generalizes to 3+ nodes Co-Authored-By: claude-flow <[email protected]>

…nced sensing) feat: ADR-074/075/076 — SNN + MinCut + CNN Spectrogram (ruvector advanced sensing)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Comparing changes

Open a pull request

Commits on Apr 3, 2026

This comparison is taking too long to generate.

Uh oh!