2022 - 2026
BSc Eng (Hons), Electrical and Electronic Engineering
SLIIT
Engineering background with work across automation, embedded systems, machine learning, computer vision, and sensor-based projects.
Machine Learning Focused Engineer
Dinupa Devinda
Machine Learning | AI/ML Developer | R&D Engineer
Machine learning focused engineer with a background in engineering, computer science, and mathematics. Experienced in data preparation, model development, evaluation, and model deployment. Interested in AI/ML work connected to real engineering and technology problems.
ML projectsComputer visionAudio classificationIMU and sensor dataEmbedded systems
ML focus
Model development, evaluation, and deployment
Computer science base
Programming, data handling, algorithms, and ML tools
Engineering base
Embedded systems, automation, electronics, and sensors
About me
Machine learning with an engineering base.
Projects across ML, software, data, and engineering systems.
Build ML workflows
Data preparation, model training, evaluation, notebooks, and deployment experiments.
Use data from real systems
Images, audio, IMU streams, sensor inputs, and embedded-device data from project work.
Build usable software around models
APIs, dashboards, databases, and software tools that help people use technical projects.
Education
Education
Engineering and physical sciences background.
2022-2026
BSc Physical Sciences - Computer Science, Pure Mathematics, and Applied Mathematics
University of Kelaniya
Computer science and double mathematics background supporting programming, algorithms, modelling, and machine learning foundations.
Projects
Selected work
Selected projects first, with the full list on the Projects page.
Portfolio chatbot/RAG and vector search
AI Portfolio Assistant with RAG and Vector Search
Portfolio chatbot that answers questions using Gemini, Cloudflare Workers, Vectorize, keyword retrieval, and clickable source links. A Dockerized local Python and Chroma workflow is included for learning and testing RAG concepts.
Search
Hybrid search
Vector search plus keyword matching.
Backend
Cloudflare Worker
API key stays server-side.
Tests
20 eval questions
Checks retrieval and unsupported claims.
AI SystemsBackendFull StackSoftware
Gemini APICloudflare WorkersCloudflare VectorizePythonChromasentence-transformersDockerNext.jsTypeScriptEvaluation harness
FYP prototype/Integrated ML and embedded system
Vehicular Black Box: Vehicle Violation Detector
Final-year vehicular black-box prototype built around Raspberry Pi hardware. It combines camera, audio, IMU, GPS/GSM inputs, violation and driving-event modules, evidence upload, backend services, and a React dashboard.
System scope
ML, sensors, backend, dashboard
Hardware
Raspberry Pi 4B
Repository
Public GitHub
Computer VisionAudio MLSensor FusionEmbedded/IoTBackendFull StackEdge AI
Raspberry Pi 4BPythonTFLiteONNXNCNNGoPostgreSQLReactTypeScript
Computer vision prototype/Computer vision and Raspberry Pi inference
Road Sign Detection with YOLO and ONNX
Two-stage road-sign recognition pipeline for Raspberry Pi. YOLO localizes candidate signs, then an ONNX classifier separates speed limits, traffic lights, no-honking signs, and other signs.
Pipeline
YOLO detector + ONNX classifier
Test set
9,062 images
Deployment
Raspberry Pi / NCNN
Computer VisionEdge AI
YOLOONNX RuntimeNCNNOpenCVPythonRaspberry Pi
Audio ML prototype/Audio ML pipeline and Raspberry Pi inference
Vehicle Horn Detection with Audio CNN
Audio classification model for detecting vehicle horn events from short microphone recordings. The pipeline converts audio into log-mel features, trains a CNN, and exports a lightweight TFLite model.
Input
1-second audio windows
Model
Log-mel CNN
Export
TensorFlow Lite
Audio MLEdge AI
TensorFlowKerasTFLitelibrosaSpecAugmentPython
FYP module/Audio and IMU fusion prototype
Crash Detection with Audio and IMU Sensor Fusion
Crash-detection module that combines audio and IMU signals instead of relying on one sensor. The project packages separate model outputs into simple fusion labels for the vehicular black-box prototype.
Signals
Audio + IMU
Models
CNN + CNN-GRU
Output
Crash confidence labels
Audio MLSensor FusionEdge AI
PyTorchTensorFlow/KerasCNN-GRUlog-melRaspberry PiPython
Published dataset and ML model/IMU feature engineering and XGBoost
Lane Change Detection from IMU Data
Recall-focused lane-change detector using BMI160 accelerometer and gyroscope data. It builds 3.5-second windows, extracts motion features, and uses XGBoost on the public BYD driving-events dataset.
Input
3.5-second IMU windows
Recall
0.9809
ROC AUC
0.9497
Sensor FusionEdge AI
XGBoostscikit-learnPythonBMI160 IMUZenodoKaggle
ML prototype/IMU time-window classification
Aggressive Driving Detection from IMU Signals
Sensor-based model for classifying normal versus aggressive driving behavior. It uses yaw-rate and acceleration windows, compares multiple models, and keeps the selected Optuna-tuned XGBoost artifact.
Input
2-second IMU windows
Model
Optuna XGBoost
ROC AUC
0.862
Sensor FusionEdge AI
XGBoostOptunascikit-learnPythonIMU windowsHugging Face dataset
Contact