ML / AI Experience

Embeddings generated for media assets - cross discipline
Centroid-based clustering - relevant for training and classification
FrameSession KV cache
Education: linear regression / forgetting curve (TesTU)
CUDA/NVIDIA acceleration (Server based, same as stack as Jetson)

Deep LLM integration
Experimental time-series text analysis - Pilot Vectors
Topic and content similarity (nearest neighbour) Image
Primarily Video, Image and Text - but all reduce to a stream
Build a Java->Python Bridge to facilitate access to python from platform

Python DSP Pipeline – Hydrophone Simulator

[ Acquisition ] → [ Spatialization ] → [ Per-Chunk DSP ] → [ Temporal Modeling ] → [ Segmentation ] → [ Persistence ] → [ Replay ]

Python-based DSP pipeline
Continuous time-series processing
Configuration-based and pluggable
Traditional DSP: RMS, stats, FFT, STFT, centroids, bandpass
ML: impulse isolation, isolation forest windowing, embedding tracking
CNN

Substantially faster than real time
Headroom for replay and experiments
Full-fidelity audio capture for later playback
Primarily: numpy, librosa, torch, scikit-learn, matplotlib
Data from web sources, unlabelled oceannetworks.ca

DAM Integration & Archival Data

Managed assets, not loose files
Elasticsearch NoSQL Metadata and Analysis - Extremely fast search
Stable identifiers throughout
Raw data preserved

Archival-grade audio storage
Storage agnostic: local, NAS, S3, ZFS
Safe reprocessing

Testing, Replay & Offline Analytics

Replay historical data
Regression on real signals

Elasticsearch for offline analysis
Compare configurations safely

Embedded & Edge Context

SignalK integration
ARM / Raspberry Pi

Real CPU and I/O constraints
Same pipeline logic