CoRE stack email archive search! And AI coding capabilities
3 views
Skip to first unread message
Aaditeshwar Seth
Mar 10, 2026, 2:27:34 PMMar 10
to learning, ac...@googlegroups.com, CoRE stack: NRM, core-st...@googlegroups.com
Hi all,
I had tried a bit of vibe coding some time back to set up some demos for the CoRE stack innovation challenge and I was thoroughly impressed with what Claude could do. These were mostly small scripts though and I was curious how far AI could go, so I took up a fun project to build a chat interface to query all the emails we keep posting on the CoRE stack googlegroup. This involved building a whole system - a web scraper to automatically click "next page" on googlegroups, download the messages, download the papers and articles linked from the messages, chunk them and get vector embeddings using small language models, upon receiving a query retrieve the matching chunks and use an LLM to generate a natural language response, and a frontend for this query interface.
I used the Cursor IDE (thanks Raman for the pointer) and within seconds it generated a whole plan. Check out the plan in https://github.com/aaditeshwar/archive-search/.cursor. Then it generated a whole lot of very high quality starter code. This didn't entirely work out of the box, but the problems were not with the logic or the overall project organization, rather with parsing annoying HTML, fixing installation and version compatibility, and lots of tweaks to handle hidden nuances of a whole lot of components that I had no prior experience of working with - ollama to run LLMs locally, mongodb for a document and vector store, selenium for scraping... In aggregate the time put into this from in between meetings and calls, it probably took me 3-4 solid days. But had I tried doing this from scratch, it would have easily taken a month or more to just learn and discover everything needed to build this.
Try it out. It's not too good! I wouldn't use it much myself, can be improved a lot with better prompts, better parsing, titling, etc. And you're welcome to contribute! But it's super impressive that all this could be done so quickly.
Everything runs on a cool GPU workstation I now have all to myself :) For comparison, time taken to generate a chat response with the LLM (qwen with 3B params, 2GB size) loaded into the GPU is < 1 sec, but with 12 cores running in parallel it takes 60-70 sec.
I do believe that the threat AI poses to a lot of coding jobs is quite real. The jobs are going to change very rapidly. Most of my work got done with Cursor's free tier but it's also not very expensive to buy (at least right now).
Adi
Co-founder, Gram Vaani; Co-founder, CoRE Stack
Reply all
Reply to author
Forward
0 new messages