Built a RAG-based support triage agent in 24 hours for a solo hackathon.
774 docs. 29 real tickets. 3 companies. One output CSV.
Here's what I built and where it broke. π§΅
I built an agentic AI that turns your resume + a job link into a brutally honest interview prep guide.
But the real learning?
Rebuilding LangChain from scratch broke how I understood agents.
Hereβs what surprised me π
#AgenticAI#LLM#LangChain#BuildInPublic
Full resume as one blob β weak, diluted matching.
Split into ~500 character chunks β sharp, focused vectors that map cleanly to JD requirements.
Chunking isnβt a detail.
Itβs the whole game.
1. Claude doesn't call your tools. Your orchestration layer does. Claude just stops with a tool_use reason and says what it needs. That distinction changes how you think about agents in general.
A company researcher, a JD scraper, and a resume matcher.
I tried to replicate what LangChain does manually. That's where everything got interesting.
Things I actually learned:
!!!! This is a scheduled tweet !!!!
exactly one year ago today i felt like i had understood how to level up and started making progress according to my standards, i hope im doing better now , hope ive been placed πππ.
Day 16/100
Lost my streak but that's how life is, getting back up, thanks to @krishij03@Wandering_Coder for exactly doing what I came to Twitter looking for.
Solved 3 more.