Ok, here is a quick list of projects (excluding hacking the computer on my telescope, as that will be difficult without being here):

Secure Router Project: Forward-port some Linux kernel patches, write a web front-end to the Quagga router, OpenFlow software and Linux Netfilter that is idiot-proof, and another that a decent user could use. I want to build something that is very, very secure and easy for Joe Average to keep secure.

Secure Router Project II: One of my objectives is to develop an open source kit equivalent to Barracuda's quarter million dollar system for securing websites. Well, as equivalent as you need to be if you're not running 250 independent servers. To do that, I need a simple, lightweight web cache, that a web server can push updates to, for static content. Squid is too heavy and has stability issues at times. It doesn't need filtering (the firewall does that) or rewrite rules (netfilter does that) but it does need to send a copy of all inbound traffic to a different process for intrusion detection.

Ground Penetrating Radar: These things are nearly useless unless used by an expert because the results are cluttered and confused. Part of the problem is that the software used assumes the ground is uniformly dry and that nothing ever refracts the signal.

Max Tegmark developed a radio telescope that could look in all directions at the same time, using FFTs and lots of receivers. My thought is that it should be possible to borrow that method to boost sensitivity and directionality. Using pulses and continuous wave, it should be possible to create a system of equations for the exact position of an object. Solve the system and you have a virtual reality map of everything underground.

Neural Networks: A neuron, in a computer simulation, has N inputs and M outputs. Each output is raised if one of a set of inputs is present. In other words, a logic gate with a truth table per output. A gate can't learn so easily, but if you have a neural net that is already trained, you just want it for recognition, then why not convert the neural net into the corresponding boolean function, optimise and then convert to a hardware language like Verilog or System C? Much faster than running it as a neural net.

Linux Autoconfig: Have a program that uses the hardware detection libraries and bus scanners to produce an initial configuration for the Linux kernel. It would greatly simplify setting it up.

Linux SELinux Autoconfig: Given a suitable test script for a program, and all the symbols and dynamically linked libraries, the idea would be to heuristically figure out the strictest permissions that the program could tolerate. Most distros provide archaic or broken configurations, because it's too difficult to set up manually. But most programs have test kits and there are test kits for Linux as a whole. It shouldn't be hard to extend this and use it to make sure everything is secure and still working.

Mandelbrot Particle Tracer: A given starting point in the Mandelbrot set will follow a trajectory. Normally, this is coloured in according to how long it takes the point to escape. Let's say, instead, you plot the trajectory itself. There are places where the point orbits a Strange Attractor. It doesn't matter which starting point you pick from the orbit, it will still orbit the same point. What you should end up with is a set of black holes, with stable orbits, n-ary orbits and areas of escape, probably all with fractal boundaries.

What if you had a program that approximated (to some arbitrary resolution) the different regions and plotted those instead? A contour map of known Strange Attractors and their gravity wells. Better yet, also provide a visual representation that allows Strange Attractors to be classified by a user.