Postdoctoral Research Fellow · The University of Sydney
AI for satellite networks, IoT and wireless communications — from routing packets across mega-constellations to drone-swarm perception and LLM agents that troubleshoot networks.
01 — About
I completed a PhD degree at the School of Engineering, RMIT University, Melbourne, Australia. This is the same institution where I completed my Bachelor's degree in Electrical/Electronics Engineering back in 2019. My research has consistently engaged with defence sector collaborations.
Currently, I am working on a defence project at The University of Sydney. Before this, at UNSW Canberra and The University of Sydney, I contributed to other defence projects on distributed AI for UAVs, specifically focusing on resource efficiency and robust object detection and tracking. Additionally, I worked on a Defence Science and Technology Group (DSTG) project at RMIT investigating AI applications for interference robustness in telecommunication systems.
My research interests span AI/ML, IoT, satellite networking and wireless communications.
02 — Research labs
Six small simulations that run live in your browser. They are simplified illustrations of the problems I work on — not the models from the papers themselves.
LLM agents can now log in to a network and debug it. SADE gives them an engineer's discipline: gather evidence layer by layer before committing to a cause. FaulT-Bench then asks a harder question — what if the ticket itself is wrong? Pick a ticket and play the agent: probe the devices, then commit to a diagnosis. Or watch a disciplined run.
A single camera loses people behind buildings. Several drones looking from different angles can fuse what they see into one ground-plane map. Drag the drones around, or switch some off, and watch the tracks on the right drop out.
A camera on a drone or a stadium roof has little bandwidth to spare. A masked autoencoder lets it transmit only some image patches and have the server fill in the rest. Choosing those patches at random throws people away; guiding the choice with a segmentation model keeps them. Raise the masking ratio, then slide the guidance down to zero and watch the players vanish.
Patches are sampled with a weight that grows with how much of a person they contain; at zero guidance every patch is equally likely. The paper's method uses a pre-trained segmentation model and a tunable power function (κ) for this, and a trained masked autoencoder for the reconstruction. Here the rebuild is a plain blur from neighbouring patches, so what is not sent is simply lost.
Laser inter-satellite links let data hop across the sky in vacuum, where light travels about 47% faster than in optical fibre. Pick two cities and watch the shortest path re-form as satellites sweep overhead. Switch the links off to see why they matter.
Computing route…
Walker-delta shell at 53° inclination, +grid links, 10° minimum elevation, time running 60× faster. Fibre delay assumes a perfectly straight cable with refractive index 1.468.
Spiking neurons stay silent until their membrane potential crosses a threshold, then fire and reset. Because they only spend energy when something happens, they suit power-starved hardware such as satellites. Raise the input current and watch the firing rate climb.
LoRa encodes each symbol as a frequency sweep — a chirp — whose starting frequency carries the data. Type a message, then drown it in noise. The classic receiver multiplies by a reverse chirp and looks for a single peak in the spectrum; it keeps working well after the signal has vanished from view.
—
Spreading factor 7, so each symbol carries 7 bits — one ASCII character here. No coding or interleaving; the two leading up-chirps are the preamble.
03 — Publications
Publications can also be downloaded from ResearchGate; refer to the corresponding publisher for the published version.
No publications match that filter.
04 — CV
The University of Sydney
CurrentUNSW Canberra
Project on UAV path planning and object detection.
The University of Sydney
The University of Sydney
Defence Innovation Network (DIN) project on distributed AI for the UAV (resource efficiency and object detection).
RMIT University, Melbourne
Defence Science and Technology Group (DSTG) project aimed at utilising AI for interference robustness in telecommunication.
RMIT University
“Smart Cities” project.
RMIT University
TopicImproving Satellite Connectivity Using Artificial Intelligence and Inter-Satellite Links
SupervisorsProf. Akram Al-Hourani (Primary), Prof. Margaret Lech, Dr. Bassel Al Homssi
RMIT University
Honours projectAutonomous Wheelchair for Non-Ambulant Children with upper limb impairment
05 — Additional materials
Introduction of the MATRIX Dataset, featuring synchronized footage from eight drones in an urban environment with comprehensive annotations for detection and tracking.
github.com/KostaDakic/MATRIX ↗
Supplementary material on our work on communication-efficient distributed multiview detection and tracking using masked autoencoders (MAEs).
kostadakic.github.io/mvmae ↗Look into our research on Low Earth Orbit (LEO) satellite networks, where we delve into the challenges in utilizing Inter-Satellite Links (ISL) for lower delays and enhanced efficiency. Our paper focuses on the critical balance between minimizing delay and conserving power in LEO satellites.
Watch on YouTube ↗My final year undergraduate project was to develop an autonomous wheelchair intended for people with upper-body disabilities such as ALS, spinal cord injuries and other neurological conditions. The outcome of this project was an autonomous wheelchair using perception sensors which is capable of taking voice commands from a user and then automatically carrying its occupant to their desired location, avoiding any obstacles along its path.
Watch on YouTube ↗