Applied Machine Learning & Computer Vision

AI, Computer Vision & Real-World Systems

A practical, production-focused track record of deploying machine learning into operational enterprise software—from embedded ONNX vision models and vector databases to custom dataset annotation tooling, vibration FFT forecasting, and agentic RAG architectures.

Facial Recognition & Vector DB FFT Vibration AI Applied AI Stack ← Back to Main Website

Engineering Mindset

Engineering Across Unfamiliar Domains

Our core differentiator is not simply knowing individual machine learning algorithms in isolation. It is the seasoned capability to enter complex, unfamiliar technical and physical environments, understand the underlying operational requirements, select or build the appropriate technology, and deliver the complete, deterministic software platform around it.

From embedded telemetry and industrial condition monitoring to retail video and high-throughput security control rooms, the constant thread has always been disciplined systems engineering. Applied AI is the natural evolution of that battle-tested toolkit.

Production Vision Pipeline

1. Facial Recognition & High-Throughput Vector Search

Architected a production-grade facial detection, recognition, and real-time alarming platform for multi-site commercial security operations (Daxicare / IpDynamics), replacing proprietary black-box camera packages that failed to meet operational accuracy and throughput requirements.

Real-Time Inference & Vector Matching Topology:

Live RTSP Camera Stream / Dahua DSS Video Snapshot 30 FPS Feed

↓

SCRFD 2.5G KPS (ONNX) — Sub-millisecond Face Detection & Landmark Extraction Native C# Model

↓

ArcFace / ResNet-100 (ONNX) — 512-Dimensional Facial Embedding Generation Feature Vector

↓

Qdrant Vector Database — Nearest Neighbor Cosine Similarity Search < 1.5ms Query SLA

↓

Score Verification Engine — Configurable Threshold Filter (≥ 82% Similarity) False Positive Gate

↓

RabbitMQ Event Dispatch ➔ Real-Time C# SOC Control Room Alarms with Mugshot Instant Dispatch

Key Technical Innovations:

Native C# ONNX Runtime Integration: Deployed lightweight ONNX models directly inside C# service processes, completely avoiding external Python server dependencies and memory leaks in 24/7 security control environments.
Vector Similarity Matching at Scale: Stored and indexed high-dimensional face embeddings in Qdrant, enabling sub-second nearest-neighbor cosine similarity queries across tens of thousands of watchlisted subjects.
Proven Operational Superiority: Delivered higher detection recall and significantly fewer false alarms than the enterprise video management system's built-in proprietary analytics.

Vehicle Identification

2. Automated License Plate Recognition (ALPR) & Vehicle OCR

Integrated automated license plate recognition into high-traffic checkpoint security workflows across 25 sites, processing active vehicle flows in real time with continuous law enforcement correlation.

Dual-Stage Localization

Utilized ONNX-based plate bounding-box localization combined with Google Tesseract OCR for reliable alphanumeric character extraction under rain, glare, and night headlights.

National Crime Hotlist REST

Connected incoming plate events via asynchronous REST APIs to the Business Against Crime national hotlist database, instantly alerting security guards when flagged vehicles entered.

Sub-Second Checkpoint Flow

Handled sustained traffic peaks exceeding one vehicle scan per second per checkpoint with zero queue backpressure and 99.8% barrier relay reliability.

Edge Video Analytics

3. Custom YOLO Tooling & Edge Video Analytics

Rather than simply consuming off-the-shelf pre-trained computer vision models, we engineer custom tooling to annotate data, retrain specialized neural networks, and deploy them directly to resource-constrained edge hardware.

Custom C# Dataset Annotation Platform

To train models on specialized domain objects (such as retail shrinkage behaviors and industrial components), we engineered a bespoke C# desktop data preparation suite:

Deconstructs raw RTSP video streams into synchronized frame sequences at configurable intervals.
Provides an ergonomic bounding-box annotation interface with automated class labeling and coordinate validation.
Automates dataset export into standard YOLO directory hierarchies with split train/val/test subsets.

End-to-End Training & Edge Deployment Lifecycle:

1. Custom C# Desktop Annotation Suite ➔ Labelled Bounding Box Datasets

↓

2. PyTorch Training Pipeline — YOLOv8 Fine-Tuning & Transfer Learning

↓

3. ONNX Quantization & TensorRT Optimization for Deterministic Edge Throughput

↓

4. Deployment to C# Control Room Systems & Android Edge Surveillance Units

Industrial Signal Processing

4. Vibration FFT Signal Processing & Predictive Maintenance (MIB)

In partnership with machinery-health specialist Simon Edmondson at CMServices Global Ltd, engineered Maintenance In Balance (MIB) — an industrial platform converting raw vibration measurements into predictive bearing failure warnings.

Signal Processing & Harmonic Forecasting Architecture:

Raw Vibration Sensor Waveforms (Imported Binary Byte Arrays from Pruftechnik)

↓

Fast Fourier Transform (FFT) Signal Processing & Spectral Decomposition

↓

Frequency Band Separation & Bearing Harmonic Characteristic Analysis

↓

Time-Series Trend Accumulation & Machine Learning Forecasting Algorithms

↓

Actionable Maintenance Intelligence: Projected Failure Window & Root-Cause Identification

Award of Excellence in Machinery Health

Sensor measurements were stored as raw binary byte arrays rather than clean machine learning tabular data. We architected the signal-processing pipeline using FFT to isolate distinct harmonic frequencies and correlate them with manufacturer bearing geometries. The system monitored over 1,000 bearings across 200 heavy industrial machines across 10 corporate clients, earning CMServices Global Ltd an industry Award of Excellence in Machinery Health.

Technical Reference

Applied AI & Computer Vision Technology Stack

Production-evaluated frameworks, inference engines, and database systems deployed in active commercial installations:

Capability / Task	Model / Architecture	Inference Engine	Deployment Target	Benchmarked SLA
Face Detection	SCRFD 2.5G KPS	Microsoft ONNX Runtime	C# Native Windows Process	< 5.0ms per frame
Facial Embedding	ArcFace / ResNet-100	ONNX Runtime	C# Microservice	< 8.0ms per crop
Vector Similarity Search	512D Cosine Distance	Qdrant Vector Database	Dockerized Linux Daemon	< 1.5ms per query
Vehicle License Plate OCR	Custom Plate Detector + Tesseract	ONNX + Tesseract OCR	C# Checkpoint Agent	< 180ms end-to-end
Custom Object Detection	YOLOv8 Small / Nano	ONNX Quantized FP16	Android Edge & C# NVR	30 FPS Real-Time
Industrial Machine Health	FFT Harmonic Decomposition	C# / Python Math Core	Web Analytics Service	Real-time streaming

Need Production AI Without Cloud Lock-In?

Let's discuss how custom computer vision, vector search, or predictive telemetry can solve your operational challenges.

Schedule Architecture Consultation Explore All 12 Case Studies