Week #08
This past week I tried to close in on what I mentioned last week — both on the code side and on the fabrication side.
I fine tuned Google's Gemini model with my own training data (explained in the previous post). I still need to add more training examples (reading all the news and giving it scores is mentally draining and so I have to space it out), but I wanted to get up and running. Even with just 80 training examples I compared the output between the default model and my fine tuned one and felt the results were closer to how I see things. This way the scoring is literally tuned to my views and makes sense.
I also quickly tested connecting to an Arduino through my server just so I know if there are any issues there. I managed to send the score from Firebase database to the Arduino using MQTT.
The only issue I did have was about a limit to usage of Shiftr.io which I used as the MQTT broker — I can probably just time the requests, but good to note this and allocate time for managing it when I get to that part.
I talked to Phil and he made me realize I don't really need to build the whole mechanism myself, but rather just buy part of it (not on the cheap side though). So I got a linear rail guide that comes built in with a stepper that makes a small stage move up and down the rail. Now I only have to make the connectors. I also started designing the enclosure in Fusion — I want to use sheet metal so I'm going that route, but I don't really know fusion and didn't want to feel restricted so I designed it in Illustrator and then used the laser cutter. I used paper as it seemed closer to sheet metal than the cardboard I had at hand.
The design is inspired by Teenage Engineering's computer case which I really like because it's simple, yet compelling. It gives this feeling I was aiming for which is somewhere between brutalistic and sleek.
The next steps are to design this in Fusion with all the components and their placements, then do a cardboard prototype (I need to figure out the sheet metal thickness). After that I'll hopefully send it ASAP to sendCutSend for cutting and bending. I think I'll do powder coating too but I'll get into that as it becomes relevant.
I updated the BOM with a couple more things Pedro suggested I'd get, like hall effect sensors for a homing functionality (figuring out where the motor is is not trivial). So far things are going according to plan but I'm mostly worried about the fabrication and how everything comes together so I'm trying to just work on everything in parallel.