A Few Ideas for Projects

I don't want to lose track of the project ideas I've been having recently, because I think many of them are really good! I also know that there isn't enough time to work on them all myself. I hope you can find inspiration from one or more of these.

Reach out to me at "contact [at] thornewolf.com" if you want to dive further into any of these.

Rapid Prototyping Python Library using GPT

Using LLMs for code generation is by no means a new idea, but solutions I've seen so far under-serve the programmer by delivering a lack-luster developer experience. The main issue with current solutions is the copy-paste loop between interacting with a chat agent to implementing their solution in your code. A secondary issue is the context switch from a separate panel/window to speak in plain language to the agent before returning to your programming domain.

Consider the DX with Github Copilot. It is strictly better than many of these ad-hoc implementations because there is less context switching between programming and conversing. Where Copilot falls short is understanding the architecture of early-stage codebases. Since Copilot primarily is an interpolator on the domain of {Your Code, Typical code with matching signatures} it struggles with broad tasks like configuring a database connection for later use, creating then using a new data type, or implementing both a function and the various functions it depends on.

Consider the following decorator syntax:

@implement_undefined
def get_user_from_db(user_id: str):
	db = get_db_connection_from_pool()
    user = db.query(User).filter(id=user_id).first()
    return user

In this example we will assume that all constructs in the implementation of get_user_from_db are undefined. I imagine @implement_undefined to catch any NameErrors and do a ad-hoc implementation & evaluation of the undefined name, passing the full function implementation as context.

I imagine that this pattern could be very powerful for rapidly prototyping backends while building out a UI.

Image Captioning SaaS

This idea is my favorite project idea that I probably will never actually end up implementing due to a lack of interest. Consider a javascript snippet that you can inject on your webpage that automatically attaches alt-text to every img tag on your page.

<script>api_key=ABC</script> 
<script src="cdn.bulk-caption-service.com/caption.js"></script> 

The actual implementation being the above modulo correct JS syntax.

Caveats and Considerations

Accessibility is a very deep field and it would be presumptuous to claim that one can hack together a genuinely useful solution for the downstream user - those with low-visibility on the web. A significant amount  of the engineering lift in this project is talking to these users and making sure your solution works before selling it to customers.

Vanity Plate Suggest & Check

Searching for a vanity plate that meets your taste preferences is a time-consuming task. You must generate a list of concepts you might want to get a plate for, generate a list of candidate words for these concepts, convert each to vanity language, order them in terms of preference, then finally manually check each one against your local DMV's database.

I imagine an LLM integration being perfect for taking out most of the friction out of this process. A user can provide the list of concept or words they are interested in and the LLM can generate a candidate set of plates, checking each one against the database proactively. Instant time savings.

Engineering Lift: probably need a one-off integration with each state's DMV. But definitely doable.

Phone Transcription and Translation Proxy Service

Combining the Twilio Voice API with OpenAI's Whisper API with some TTS model could prove to be a very convincing tech-demo for real-time phone call transcription & translation. This could unlock all sorts of cross-locale business communication points of friction and frankly might be a (slightly) more moral approach to the phone call bot marketing spam we are currently being subjected to.

In a minimal use-case we could see this being used by any individual who would benefit from getting a transcription from their phone calls.

Amazon MTurk on Discord

MTurk is a platform that has driven massive amounts of value for any process that needs a lot of data to work - from university research to Industry ML labeling. I think there is an opportunity to introduce a new-age version of MTurk where tasks are delivered vis Discord, bringins a well defined interface to an old problem.

MTurk seems to drive ~$6M/year of revenue and ScaleAI $290M (both according to Google). This means the total addressable market is large but involves a fight against VC.

Conclusion

I think each of these projects have pretty clear paths between zero -> tech demo -> MVP -> product such that I would be excited to work on any/all of them. But time is scarce, and attention scarcer. I hope that someone can find inspiration from these and if you do please reach out!

contact [at] thornewolf.com