Pradyupdates first email
What?
(Inspired from Judah, Trevor and Alexey Guzey )
At first I was going to write some teary-eyed content about how its so nice to be trusted by so many people on the internet and y'all root for me and care for me. But I'll skip that and get to the point:
- I'll give you updates every month to three months, you'll (hopefully) read them and even more hopefully tell me what you think of them
- These updates will be about
- My projects (programming projects, blog etc.)
- My other ambitions (doing art on OneNote)
- Academics (although I hope to not talk about it here)
- random links
- professional life (internship search, job search etc etc)
- random ramblings
- other personal life updates
I'll be as open as I can to a group of ~20 people, and redact whatever i need to. But you should get a decent picture of my life from this. also if you want to unsubsscribe, lmk! I won't take it personally.
so let's go
Projects!
Automated Singapore data analysis bot
The big deal of the last few months was me trying, failing, retrying, refailing (repeat once or twice again) to make a RAG app for Singapore's public data. I took it down (or rather stopped running a cron job that kept the Render backend from being shut down) because I'm running out of Render free hours. The demo video is here, and the site is here. It's going to be ridiculously slow because Render's cold start time is like 7 minutes versus the advertised 50 seconds.
the workflow is that:
1. get dataset (either user selected one or impute using HF 0 shot model)
2. give dataset schema, head, and any other details as part of prompt to GPT-4o
3. Ask GPT-4o to write code to answer the user query
4. Send answer (from the code) to the frontend
5. Render chart from the selected/imputed dataset
This hides a lot of complexity and design choices, but its a simple explanation of what to do.
Now what?
I think its pretty dumb / a stroke of luck that the AI agents world has not (yet) found its way to the world of economic data analysis. What does that mean? There's lots of money spent in making charts and graphs in economic "research" reports, and equity "research" reports. These are usually simple enough that a LLM can write code to make one, but the main problem lies in finding the correct dataset (search), writing the correct prompt, and doing graphing correctly
My goal is simple: build the world's first automated economic data analysis assistant. (Sidenote: such a shame that the word copilot is exclusively used by Microsoft! It describes exactly what this is, but if I name it that, people will probably confuse it with Microsoft's things?). I'm looking for a name for this. The current one: EconDataGPT is quite terrible
This is my main priority over the next 6 to 8 weeks. The first step in this is to add search over FRED using their API, and then start talking to people about it. If you know people working in macroeconomic research (the kinds who write reports about the next 2 quarter's GDP growth) please let me know how to contact them.
Branching Ideas GPT
Continuing on the theme of GPT wrappers, me and Lavanya built an idea generator with branching. It's basically a better ChatGPT interface, which allows you to go down different rabbitholes with your ideas, and then if you don't like it, go back to the start and create a new branch
I'm quite surprised that AI labs haven't come up with this already, but it's a good one. We'll make some improvements to this, but have been stopped in the recent past by a combination of other work, procrastination and laziness
Upcoming projects
the main priority is the data analysis bot. I want to sell it to people who use it so I can be free of the need to do leetcode and other code monkey work
A small project im working on is a terminal based life system for myself. Its not very complicated, but the main point is - I want to extend existing todo apps in minor ways and can't do them (or run into issues with pricing) - I don't want ANYTHING on the browser. The browser usually ends up in wikipedia binge sessions or twitter or any other thing not conducive to productivity - Simple writing it down in a notepad file / markdown / physical paper is insufficient because I want things like recurring tasks, tagging, reminders etc.
So, the only choice left is to build my own. It's a CLI app, and the main purpose is to be a todo-list/project tracker/time logger all in one for me. I've been told I should use org mode but I have my reasons (don't want to use LISP, want it stored in a DB so I can build more complex apps).
Blog
The blog will remain archived as it is now. I have a review of Jeff Ding's Technology and the Rise of Great Powers coming up in a month or so.
What next?
By the end of the next quarter, I need to 1. get a working version of the app done 2. Apply to the school of computing venture initiation program 3. DM a shit ton of people about it 4. Do a lot of leetcode problems 5. Apply to internships (because i want some cash and credentials) 6. get a better format for this newsletter
Perhaps the lesson to take for me is from this is to understand that I have to do everything to win. And I'm not at the moment, doing everything.