Engineering Portfolio

SonikSearch

AI-powered audio sample library manager with on-device inference, vector search, and cross-platform native deployment. Solo-built. Commercially shipped. Revenue generating.

C++ ONNX Runtime Core ML DirectML Qt SQLite HNSW Shipped Product Solo Engineer

SonikSearch is a professional sample library manager for music producers and sound designers. It lets users browse, search, analyze, and audition audio sample collections — from 10,000 files to millions — as a standalone desktop app or as a plugin running directly inside a digital audio workstation.

The product is commercially available at $89, generates revenue, and has been covered by major industry publications including KVR Audio, Rekkerd, and The Beat Community. I designed, built, and shipped the entire product as a solo engineer — architecture, inference pipeline, UI, CI/CD, cross-platform deployment, and ongoing maintenance.

On-Device ML Inference Pipeline

  • Integrated a CLAP (Contrastive Language-Audio Pretraining) model to generate 1024-dimensional float32 embeddings from audio samples, enabling semantic search — users type natural language descriptions like "warm pad" or "punchy kick" and find matching sounds regardless of filename.
  • All inference runs locally on the user's machine. No cloud, no network dependency, no data leaving the device.
  • Built on ONNX Runtime for cross-platform model execution. On macOS, uses the Core ML execution provider to accelerate inference on Apple Neural Engine (ANE) hardware. On Windows, uses the DirectML execution provider for GPU-accelerated inference.
  • Designed for large-scale imports with background processing so the library remains usable during analysis.

Vector Search at Scale

  • Built an HNSW (Hierarchical Navigable Small World) vector index for approximate nearest-neighbor search over audio embeddings, enabling the "Find Similar Sounds" feature.
  • Embeddings are quantized to int8 for the HNSW index, reducing memory footprint and improving search throughput while maintaining search quality.
  • Designed the indexing architecture to scale to millions of samples with sub-second query times.
  • Users can adjust similarity weighting across nine acoustic dimensions (brightness, warmth, harmonicity, attack, pitch, and more), giving fine-grained control over what "similar" means.

Audio Analysis Engine

  • Automated extraction of BPM, musical key, loop/one-shot classification, and category tagging across 31 subcategories.
  • Computes detailed audio characteristics: brightness, warmth, crest factor, harmonicity, attack profile, RMS energy, and duration.
  • Analysis runs in a background task queue — the library remains browsable and searchable while imports process. Users can pause, cancel, and reorder tasks.

Cross-Platform Desktop Application

  • Built in C++ with Qt. Ships as a standalone app and as a plugin in three formats: VST3 (cross-platform), Audio Unit (macOS), and AAX (Pro Tools / macOS).
  • Supports macOS 12+ (Intel and Apple Silicon) and Windows 10/11 (x64 and ARM).
  • SQLite database backend for metadata storage. Designed for libraries ranging from thousands to millions of samples.
  • Real-time audio playback with timestretching and key transposition. Ableton Link support in standalone mode.

Shipping & Operations

  • Tag-based CI/CD pipeline for automated builds and releases across both platforms.
  • AAX plugin signing via PACE/iLok integrated into GitHub Actions (cross-platform signing architecture).
  • Handles commercial licensing, customer support, product updates, and pricing (including regional PPP pricing for international users).
LanguageC++
UI / Plugin FrameworkQt
ML InferenceONNX Runtime · Core ML EP (macOS / ANE) · DirectML EP (Windows / GPU)
Embedding ModelCLAP (Contrastive Language-Audio Pretraining)
Vector SearchHNSW with int8 quantized embeddings
DatabaseSQLite
Build / CIGitHub Actions · tag-based release pipeline
Plugin SigningPACE / iLok (AAX) · Apple codesigning · Windows Authenticode
PlatformsmacOS (Intel + Apple Silicon) · Windows (x64 + ARM)
Plugin FormatsVST3 · Audio Unit · AAX

The domain is audio, but the engineering problems are general.

On-Device ML Inference

Deploying a neural network to run efficiently on consumer hardware with no cloud dependency — managing runtime selection, hardware acceleration (Core ML / ANE on macOS, DirectML / GPU on Windows), and embedding quantization.

Vector Search at Scale

Building and querying a high-dimensional vector index over millions of items with tunable similarity metrics and int8 quantization for memory efficiency.

Cross-Platform Native Dev

Shipping a C++ application across two operating systems, multiple CPU architectures, and three plugin format APIs — each with its own hosting environment, lifecycle, and compatibility requirements.

Solo Product Ownership

Architecture, implementation, testing, CI/CD, release management, customer support, and commercial operations — end to end.