// portfolio

Production ML Projects

Real systems, real impact. Every project deployed in a production environment with measurable results.

6+

Production Systems

~30%

Safety Improvement

40%

Pipeline Reliability

<50ms

Inference Latency

Computer Vision🚛

Fleet Driver Fatigue Detection

Real-time driver drowsiness detection system using MediaPipe face mesh + LSTM temporal classification. Processes RTSP camera feeds and sends live MQTT alerts to fleet operators when fatigue signs are detected. Deployed across an active commercial fleet with continuous monitoring.

~30% reduction in fatigue incidents
PythonMediaPipeLSTMMQTTRTSPRaspberry PiOpenCV
// private repo
Edge AI⛏️

Mining Anomaly Detection Engine

On-edge TensorRT inference engine deployed on Raspberry Pi for real-time anomaly detection in mining operations. YOLO-based object detection + UNet segmentation for hazard zone identification without cloud dependency. Achieves sub-50ms inference latency.

Sub-50ms edge inference latency
TensorRTYOLO v8UNetRaspberry PiPythonRTSP
// private repo
Data Engineering🏗️

ETL Reliability Platform

Airflow-orchestrated ETL system with DVC data versioning, Feast feature store integration, and Kafka-based event streaming. Includes automated lineage tracking, retry logic, and Slack alerting. Reduced pipeline failure rate by 40% across multi-source data ingestion pipelines.

40% fewer pipeline failures
AirflowDVCFeastKafkaDockerAWS S3Python
// private repo
Data Visualization📊

Fleet Safety Analytics Dashboard

Streamlit-based operational dashboard for fleet safety teams. Real-time GPS/Haversine route visualization, driver anomaly heatmaps, incident timeline, and configurable alert thresholds. Connects to live Kafka telemetry stream and MLflow model registry.

Real-time fleet monitoring for ops teams
StreamlitGPS/HaversineKafkaMLflowPythonPlotly
// private repo
Edge AI⚡

Edge ML Inference Engine

Modular, plug-and-play TensorRT inference framework for deploying PyTorch models on Raspberry Pi. Handles model conversion, INT8/FP16 quantization, batching, and streaming inference. Designed for industrial edge deployment without cloud dependency.

Modular edge deployment framework
TensorRTPyTorchRaspberry PiPythonUbuntuONNX
NLP📝

NLP Safety Report Classifier

BERT-based multi-label text classifier for mining safety incident reports. Auto-tags hazard categories (electrical, structural, chemical, human error) with confidence scores. Integrated into incident management workflow, cutting manual triaging time by ~60%.

~60% reduction in manual triaging time
BERTHuggingFacePyTorchscikit-learnFastAPIDocker
// private repo

Most of these systems are built for private enterprise clients and deployed under NDA. Code samples and technical deep-dives are available upon request — reach out via the contact page.