Thasanka Kandage

Software Developer Intern

Ericsson

• Built a multi-agent AI system with LangGraph and FastAPI, wiring in Jira, Trouble Reports, Confluence, Gerrit, and Jenkins to automate build workflows and engineering troubleshooting. 1,000+ conversations recorded.
• Cut retrieval latency from ~20s to under 1s across 3M+ records by fusing HNSW vector search with PostgreSQL keyword search using reciprocal rank fusion.
• Built an ingestion pipeline that chunks and embeds ~100 Jira tickets a week into a PostgreSQL/pgvector knowledge base, replacing manual updates.
• Containerized the TypeScript frontend and FastAPI backend, built GitLab CI/CD pipelines, and managed Kubernetes manifests covering Ingress, RBAC, secrets, and persistent storage.
• Automated downlink/uplink nodal time-alignment testing with a Python SCPI/PyVISA control layer for oscilloscopes, signal generators, analyzers, and CPRI capture over raw sockets.

AI Engineer Intern

Customer Maps

• Built an Express.js backend on Cloud Run powering a Shopify AI agent system, connecting Shopify APIs, Vertex AI agents, and MCP servers to automate store operations.
• Improved product search relevance and recommendation consistency with prompt engineering, RAG, catalog embeddings, and context management.

Research Assistant — UAVs & Path Planning

University of Alberta - Dr. Nima Shirzad-Ghaleroudkhani

• Developed an autonomous survey routing system that cut drone flight time by 46% while holding 91% structure coverage.
• Applied 3D A* search and nearest-neighbor optimization to generate collision-free waypoint paths around a bridge pillar.

Research Assistant — Thermal Energy Storage

University of Alberta - Renewable Thermal Lab - Dr. Taha Manzoor

• Ran multi-objective optimization of a thermal energy storage system, tuning five design parameters with genetic algorithms in MATLAB. The work became a first-author conference paper.
• Contributed to two PyTorch models: a physics-guided neural network with embedded physical constraints, and an LSTM for forecasting storage tank behaviour.

Machine Learning Engineer Intern

Hub for Neuroengineering Solutions

• Reached 93% mAP@50 on mice imagery by training CNN and Vision Transformer models for object detection and pose estimation.
• Cut manual labeling time with an active-learning annotation tool that retrained on user corrections and surfaced low-confidence samples for review.
• Improved dataset diversity using t-SNE, DBSCAN, and convex-hull analysis.

Software Team Co-Lead

UAlberta RoboMaster

• Contributed to a C++ real-time auto-aiming system using OpenCV target detection and PnP pose solving, and built the Python camera calibration tooling behind it.
• Placed 9th of 20 in 3v3 at RoboMaster North America 2025 in San Diego.

UFC Fight Dataset & Predictor — (Check It Out)←

Scraper for every fighter, event, and fight stat on the official UFC site. Published the cleaned dataset on Kaggle — 1,500+ views, 300+ downloads. Random Forest, SVM, and XGBoost models on top of the engineered features.

TKShell — custom UNIX shell

Piping, I/O redirection, background jobs, and built-ins, with signal handling so child processes can't take the shell down with them.

AI Sudoku Solver — (Try It Out)←

Reads a Sudoku from an image and solves it. CNN digit recognition after an OpenCV pipeline of contour detection, thresholding, and perspective warping, then recursive backtracking.