+ Hobby Projects
+
+ Over the years I've hacked away at various personal projects. More
+ recently I've experimented with . My preference is always to build,
+ run and host applications locally, which includes this page!
+
+
+ Recently my interesets have shifted slightly to large machine learning
+ models, and have messed around with Stable Diffusion (mainly with
+ Invoke AI) and Llama
+ language models in
+ rustformers. The
+ latter has been a bit dissappointing (at least for its programming
+ ability), and I think my job will stick around a while longer!
+
+
+ Finally I've thoroughly enjoyed writing in Rust, mainly the
+ efficiency, ease of use and correctness that come from using this
+ programming language. One example was in the
+ FastCoster
+ project, where I reduced the time taken for processing some demo data
+ on the costing product from ~1.5 hours to ~7 seconds on a
+ laptop/desktop, or ~36 seconds on a smartphone. This was mainly due to
+ not using SQL Server, and using a custom algorithm in overhead
+ allocation that significantly reduced memory consumption and the
+ number of required calculations.
+
+
+ PiCar
+ Source
+
+
+ This project originally involved communication between a Raspberry Pi
+ and a Traxxas Slash using the Pi's GPIO to control the steering and
+ throttle of the RC Car. This was mounted on some 3D printed
+ brackets.The steering and throttle are set using an iPhone/Android
+ application connected over WiFi.
+
+ Over time this worked as a base to explore other ideas, namely:
+
+ -
+ SLAM: Using BreezySLAM and a 2D RP Lidar A1, the Pi can map out an
+ area and send this information to the controlling phone.
+
+ -
+ Depth Prediction: Using the Pi's camera and an Intel Neural Compute
+ Stick (NCS), the Pi could process camera data and use a custom
+ implementation of the
+ FastDepth
+ Neural Network to add 3D sensing capabilities.
+
+
+
+ Recently there have been efforts to port the backend to Rust, with the
+ 2D Lidar sensing and control completed. The Python BreezySLAM
+ implementation is currently unfinished, mainly due to distractions
+ from other projects
+
+
+ Depth Prediction
+ Source
+
+
+ From the PiCar project, I explored many different implementations of
+ monocular depth Prediction and 3D SLAM solutions, including
+ implementing my own algorithms and trainers for depth prediction that
+ perform well on constrained devices.
+
+
+ This gave me a solid foundation on Tensorflow/Keras and computer
+ vision. It also helped expand my knowledge on machine learning from
+ university/online study, as I previously had not explored models this
+ large, or specifically computer vision related models.
+
+
+