Ovinda Sakun.

Built during my Sri Lanka Telecom internship. The pipeline combines YOLOv8 ONNX detection, Vertex AI Gemini OCR and Sri Lankan plate formatting logic with MySQL logging, ESP32-driven LED displays and a CustomTkinter dashboard.

My contribution

Engineered the recognition pipeline and integrated software, database logging and hardware control.

My final-year research project integrates an AS608 fingerprint sensor with ESP32 verification, an AES-128 Windows host loader and machine learning agents for Data Loss Prevention and Host-based Intrusion Detection.

My contribution

Engineered the portable workspace and integrated biometric authentication with host-side security monitoring.

Developed at Sri Lanka Telecom using CCTV input and YOLOv8 inference on Raspberry Pi 4. A fault-tolerant ESP32 LoRa mesh detects node failure and dynamically reroutes alert messages.

My contribution

Developed the computer vision workflow and integrated the edge detection system with mesh-based alert delivery.

Architected and deployed an observability environment on an Ubuntu VM using Prometheus, Node Exporter, Alertmanager and Grafana Loki for Network Operations Center telemetry.

My contribution

Deployed the monitoring stack and connected system metrics, log collection and alerting.

The app preprocesses uploaded audio, saves voice profiles with reference WAV files, and uses XTTS-v2 zero-shot synthesis. Processing trims silence, normalizes audio and exports 24 kHz mono WAV references.

The neural model loads on the first generation request and remains cached. Generation runs through FastAPI’s threadpool while the interface stays responsive. Clips, profiles and generated audio stay local; synthesis uses CPU inference.

Practical constraints

Generation can take 60–90 seconds or longer on low-spec dual-core laptops. Initial setup requires downloading the model.

The modular pipeline generates anchor speech with Microsoft Edge TTS or accepts existing audio. A local HuBERT model extracts features, an optional FAISS index retrieves voice characteristics, and a VITS-based generator produces the converted audio.

The CustomTkinter interface supports hot-swapping .pth models and .index files, pitch shift, index rate and F0 extraction settings. Output can be played in the interface or exported as WAV.

Implementation

CPU-based inference for compatibility, with output sample rates derived from the selected model. Voice conversion runs locally; text-to-speech anchor generation uses the online Edge TTS service.