Hasarindu Perera

Software Engineer (AI) at Vega Innovations

• Project lead for NOVA, an autonomous security quadruped I conceived: leading a team of four engineers and owning the system design, roadmap and delivery.
• Deployed Autoware on NVIDIA AGX Orin with a Hesai XT16 LiDAR (Nebula driver).
• Built a monocular depth estimation pipeline (relative and metric) to improve point clouds for detection and mapping.
• Built a bird's-eye-view system fusing LiDAR depth, clustering, speed estimation and camera pose.
• Quadruped: LiDAR obstacle detection in a custom ROS package, so the robot stops on its own near obstacles.

AI & Robotics Instructor (part-time) at ANKA Technologies

Classes and workshops on AI, machine learning and robotics with Python, PyTorch, TensorFlow and ROS. I guide student projects in computer vision and robot navigation.

Associate AI Engineer at Vega Innovations

• Multi-agent system for AV simulation (FLAMEGPU2, UniAD, DaveNet) and an end-to-end driving pipeline (ResNet, RegNet, BiFPN).
• Python sensor-fusion module for LiDAR, RADAR, camera, GPS and CAN data.
• Go2 quadruped: Go2 SDK + ROS 2 integration, SLAM localization with Nav2, RL control trained in Isaac Gym.
• FaceRec: moved the system from Samsung tablets to Rockchip boards, cutting hardware cost per device by about 80%.
• VelloPass (licence plate recognition): image enhancement research and CI/CD.

Intern AI Engineer at Vega Innovations

• Implemented TRPO, PPO, CPO and PCPO from scratch for safe RL in autonomous driving; experiments in CARLA and MuJoCo with TorchRL, CleanRL and OmniSafe.
• Face recognition attendance system on Raspberry Pi 4B, with Intel RealSense depth anti-spoofing.
• Camera-to-BEV semantic segmentation with UNetXST (85% accuracy).

Mentee, AWS AI & ML Scholarship at Intel

Top 500 cohort. Mentored by Ali Aram, Platform Application Engineer at Intel.

Contributor at Omdena

Autism Early Detection Project with Dr. Shaista Hussain: dataset collection, literature review and experiments on neuroimaging (ABIDE I & II, fMRI and structural MRI).

Founder & Director at heartivVe (Pvt) Ltd

Built 20+ business websites and full-stack products, plus CMS-based learning platforms, mostly in Node.js and JavaScript.

NOVA: an autonomous security quadruped

NOVA is a four-legged security robot that patrols a site on its own, notices what has changed, and lets operators watch and talk through it from anywhere. I came up with the idea, named it, and led it from a Unitree Go2 prototype to a product. The first prototype is complete and working, and NOVA is now ready to pitch to clients.

• Led a team of four engineers across perception, speech, navigation and the operator platform, and owned the system design, roadmap and delivery.
• Perception on the robot: detecting and tracking people (YOLOv8, ByteTrack), recognising authorised faces, and spotting what has changed along a patrol route from LiDAR scans.
• Autonomy with ROS 2 on an NVIDIA Jetson: mapping buildings it hasn't seen before, following patrol routes, and returning to its dock when the battery runs low.
• An on-device voice assistant running a local LLM, so staff can send the robot somewhere just by speaking. Operators can also talk through it and sound a warning siren.
• The operator side: a React web dashboard with live video, a map and remote control, a Flutter mobile app with push alerts, and a terminal app for field work. Patrol logs sync to the cloud.
• Evaluated newer approaches, such as world models and vision-language-action models, against proven ones, and chose what was reliable enough for production.

Autonomous vehicle

Vega's self-driving car research platform. I've worked on it from my internship onwards.

• Researched safe reinforcement learning for driving, implementing TRPO, PPO, CPO and PCPO from scratch and testing them in CARLA, MuJoCo, Safety Gym and Highway-Env.
• Built a multi-agent system for AV simulation, and an end-to-end driving pipeline from data collection to evaluation.
• Wrote a Python sensor-fusion module that brings LiDAR, radar, camera, GPS and CAN data into one place for perception work.
• On the current car: set up Autoware with a LiDAR, built monocular depth estimation (about 5% error in controlled tests) and a bird's-eye-view perception system.

Go2 quadruped

Navigation, localisation and control for a Unitree Go2 quadruped. This work became the starting point for NOVA.

• Integrated the Go2 SDK with ROS 2 for motion planning and locomotion control.
• Set up SLAM-based localisation and navigation with Nav2 and LiDAR.
• Trained reinforcement learning controllers in simulation (Isaac Gym).
• Built LiDAR obstacle detection as a ROS package, so the robot stops by itself when something is in the way.

FaceRec

A face-recognition attendance system that can't be fooled by a photo or a screen, built to run on low-cost hardware.

• Moved the system from Samsung tablets to Rockchip Linaro boards, cutting hardware cost per device by about 80%.
• Researched anti-spoofing across wavelets, colour disparity, thermal sensing, depth sensing and ensemble methods, reaching about 95% spoof detection.
• Built and tested it on Raspberry Pi, Rockchip boards and NVIDIA Jetson Nano.
• The first version used an Intel RealSense depth camera and facial landmarks (open-source part).

Bird's-eye view

Research alongside the autonomous vehicle project: turning ordinary camera images into a semantically segmented top-down view of the scene. I implemented a UNetXST network that reached 85% segmentation accuracy.

VelloPass: licence plate recognition

An AI system that reads vehicle number plates. I researched image enhancement so plates are read reliably in poor light and bad weather, and helped build and maintain its CI/CD pipeline for testing and deployment.

VegaGPT

An internal chatbot for knowledge sharing in the AI team, giving contextual answers from the team's own material. Built with LangChain and OpenAI models.

ChargeNET OS

A slimmed-down custom Linux operating system for Vega's EV charging platform. It uses far less memory and CPU than before, and cut storage from 64 GB to 8 GB, with a 2 GB portable image.

QuPrep

An open-source Python library that prepares classical data for quantum machine learning. It cleans and encodes datasets using 15 encoding methods and exports ready-to-run circuits to 8 frameworks, including Qiskit, PennyLane, Cirq and TKET. Published on PyPI (pip install quprep) with a citable Zenodo DOI, and listed in QOSF's awesome-quantum-software.

QuantumEdge: Global Industry Challenge 2026

Phase 3 team submission to the qBraid, MITRE and JonesTrading Global Industry Challenge. Our team of four used quantum reservoir computing to forecast financial market volatility, comparing it against classical baselines and validating on IBM quantum hardware. I built the classical baseline pipeline and led the data and machine-learning work.

QuantumStuff

A free, open-source place to learn quantum computing, from beginner tutorials and interactive activities to variational algorithms and error correction. It includes c.js, a browser-based logic-circuit simulator I wrote.

qops

An open-source tool for tracking, comparing and reproducing quantum computing experiments across Qiskit, PennyLane and Cirq, with a local dashboard and command-line tools.