Writing · 10 posts
Writing
Notes from building an evaluated climate-risk agent in public. Most posts here were first published on Hashnode and keep their original date.
The eval set that graded itself
A held-out set is only doing its job if first exposure costs you something.
My agent refused to say Delhi is warming
Non-stationary GEV, likelihood-ratio tests, and the engineering discipline of refusing to hallucinate climate trends.
My agent learned to read
The plan step was an if-statement over a hardcoded location. Now a typed pipeline parses, geocodes and maps any question to an IPCC region, or refuses.
I made my AI agent tell me exactly what every answer costs
What a single answer costs, and whether the service holds up under a real HTTP request instead of a notebook cell.
When my evaluation reached zero errors, I hired a stricter judge
A claim-level judge, added next to the citation checker, found the same table-chunking bug a second time. Fixing it took both error cells to zero.
How a chunking bug taught me more than a model failure
Three wrong refusals, traced to a table chunker that kept each row and lost the column header saying which warming level a number belonged to.
My favorite number in this project is zero. Twice.
How a RAG system earns numbers you can defend: the evaluation came before the retriever.
My hazard numbers passed every test and were still wrong
Two weeks into this project, my code told me the 100-year heat extreme for Rourkela, an industrial city in Odisha, India, is 38.1 °C.
Building a climate-risk agent in public, week 1: why I built an agent, not a chatbot
I'm building a climate-risk analyst agent out in the open, one week at a time.
5 mistakes that cost me GSoC and LFX
It took me five rejections to understand this: I had the merged pull requests.
Contact / Rourkela, Odisha, India