Startup ideas
001: AI workbench to version and manage complex prompts
Managing, versioning, and evaluating complex multi-stage prompt pipelines is a pain. It's easier to run a single prompt in any API front-end but it's a pain to export, eval, and re-import into another prompt within the same workbench.
Tons of friction.
002: A simple chrome extension that summarizes an AI chat
Helps with exporting chat one AI model to use it in another. Similar to the 'switch chat' feature in Cursor that's used when the context gets too long.
003: Bloodwork scanner and interpreter
Current bloodwork results are difficult to understand without medical guidance. Plus, the provided 'ranges' are based on averages for the population, not what's optimal for a given demographic.
004: An easier way to sell secondhand stuff on marketplaces
Built a proof-of-concept for this that lets you scan items around the house and gives you title, description, and a price estimate based on your location. This helped us sell a bunch of stuff before moving back to the UK. Things that help people make money tend to work. This could easily be a 'feature' not a 'product', but the benefit would be to sell stuff across multiple different marketplaces instead of being locked into one.
005: Realtime walking or driving tour with AI-guides
We used an app with pre-recorded guides as part of a tour around the Cathedral on Gran Canaria during our honeymoon. It was great. I'd love to have something like this to explain the history of a city while I walk around it. This could use phone GPS to know where you are and what's around you - and adding in realtime speech can turn this into a dialogue instead of a one-way conversation.
006: AI tutor for software companies
Most enterprise software is complex and is confusing to use. Enablement and L&D for users usually sucks. People don't learn anything from passively watching videos or skimming over documents. The real learning comes when a user has a real problem and solves it for themselves. Embedded guidance within the app that has the context of where a user is, what they're trying to achieve, and broader information about the account can be used to guide them. It's like getting a tailored recipe to build a cake.
For now, the user would still be doing the tasks, but eventually agents should be able to make the task list and execute.
007: Timestamp and geo-locate photos for construction
Improve async work across worksites by allowing builders to share photos and progress of work. This is starting to happen, but for a manager it's tricky to track progress and ensure work is being done on time. A tamper-proof photo with embedded watermarks can help increase