Aswani Sahoo

ML/AI engineer. I build LLM agents, RAG and evals in public. 34 merged PRs: PyTorch, ExecuTorch, graph_weather, krkn-chaos, MalariaGEN. NIT Rourkela '27

[Aswani Sahoo @AswaniSahoo2]

@pranavgawai_ @jackedAJ jsut go with him. Not promotions

[DeepSeek Harness @DeepSeekHarness]

https://t.co/gwwIHtyCRQ

[Aswani Sahoo @AswaniSahoo2]

I wrote both up on the issue and asked two questions instead of proposing a plan: one mesh or two? Two-way coupling or one-way?

Part 2: how short replies from the maintainer steered 16 merged PRs.

The issue:
https://t.co/AM04pBGKUk

[Aswani Sahoo @AswaniSahoo2]

Issue #3 in graph_weather was opened in Feb 2022. When I picked it up in May 2026, it was still open.

Part 1 of 5 on my open-source weather ML work: what I read before writing a line of code.

[Aswani Sahoo @AswaniSahoo2]

In the code: to let a region move, the encoder's graph has to rebuild whenever the region's coordinates change between batches.

Two designs: one mesh refined over the region, or a second regional model fed by the global one at its edges.

[Aswani Sahoo @AswaniSahoo2]

Full write-up, with the extraction code and the scope-guard fix:
https://t.co/Z9vuwIgvMY

[Aswani Sahoo @AswaniSahoo2]

Two AI subagents burned 250,000+ tokens each trying to write my benchmark. They banked zero questions.

First Principles Log #09 of my climate-risk agent, built in public:

[Aswani Sahoo @AswaniSahoo2]

After the fixes: 0 false answers, and Recall@3 at 87% on the 70 answerable questions.

Still refused: 21 of the 70 it should answer.

How do you chase false refusals without tuning to your held-out set?

[Aswani Sahoo @AswaniSahoo2]

@thtbee_ @RyanSael Yes it's cool but I don't have windows to try it as I am in Linux 😕

[Aswani Sahoo @AswaniSahoo2]

@Sharky_XBT @LucaNetz 😕😕 what the .. seriously !

[Aswani Sahoo @AswaniSahoo2]

Built SplitSmart for #KiroUniversity: a group expense splitter usable from the browser or any AI agent via its own MCP server. Specs, hooks, PBT, powers and custom agents.

Repo: https://t.co/esw3oY5fCL
Demo: https://t.co/Go8rrmhSTv

@kirodotdev #BuildWithKiro

[Aswani Sahoo @AswaniSahoo2]

The unglamorous fix: 300 parametric bootstrap refits took 16.0s cold.
Warm-starting Nelder-Mead from the parent MLE parameters dropped it to 10.5s in 10 lines of Python.

Full post + repo:
Blog: https://t.co/1dboUya6FK
Code: https://t.co/o9G6GsXnoD

[Aswani Sahoo @AswaniSahoo2]

Berlin: +0.76 °C / decade warming (p < 0.0001).
Delhi: p = 0.56.

Ran 60 years of daily extremes through our climate agent.
It switched Berlin to a drifting trend, but refused to claim Delhi is warming.

Why statistical refusal is our most honest engineering choice:

[Aswani Sahoo @AswaniSahoo2]

Why did Delhi return p = 0.56?

In South Asia, regional aerosol pollution dampens observational surface warming in daily station records.

The code didn't know atmospheric physics. The likelihood-ratio test simply honored the data instead of hallucinating a narrative.