I've worked on a lot of things. Many of them failed. Here's a list.
- A fundraising platform for high school sports teams. The neat idea here was that coaches could share key stats and highlights from each game and allow donors to pledge a certain amount of money according to a game statistic (ie. $1 per touchdown). Unfortunately, most teams fundraise early on in the season to handle upfront costs like equipment, so it wasn't always feasible for teams to get their funding over the course of the season.
- Nature has all kinds of nifty tricks up her sleeve, and there's a whole field known as biomimicry founded on understanding these tricks and repurposing them for other problems. We were trying to mimic the water collection properties of spider silk to build a material capable of hyper-efficient dehumidification. Currently, HVAC systems bring in outdoor air and dehumidify it by overcooling it (which essentially squeezes water out of the air) and then heating it back up to a more comfortable indoor temperature. We imagined placing humidity "filters" in HVAC systems instead of wasting energy overcooling and reheating air. We were able to design materials fairly well to control water transport (aggregated micro water droplets into larger water droplets which could roll of the surface into a collection pipe), but nucleation (converting water from a gas in the air to a liquid on the surface of our screen) proved more challenging.
- Clothes are designed to fit the average person, that's how you sell the highest volume of clothes. That kinda sucks though if you're not average and you're left with clothes that don't quite fit you just right. Our whole idea here was to allow customers to take a 3D scan of their body and algorithmically generate patterns for clothing which would fit their body perfectly. The 3D laser scanners we used were expensive and hard to scale, although with advances in computer vision this might become feasible with simple cameras (which are cheap!). I'd love to revisit this one day.
- Financial exchanges use something called an orderbook to keep track of buyers and sellers, and these orderbooks are used to match interested parties and facilitate a trade between a buyer and seller. Sometimes these books are imbalanced where you have more interested seller than buyers, or vice versa. We worked on building machine learning models which could forecast short term price fluctuations based on this orderbook data, which would then be used for algorithmic trading. It turns out financial data can be really noisy. I wrote about this experience in more detail here.
I'm sure there will be more to add in the future :)