Aswani Kumar Sahoo. ML / AI engineer building and evaluating LLM agents, RAG and MCP tools. Final-year B.Tech at NIT Rourkela, graduating May 2027.
Portfolio: aswanisahoo.github.io · Résumé (PDF) · LinkedIn · X · Email
| Project | What it is | Result |
|---|---|---|
| Climate-Risk Agent (live) | Five-node LangGraph agent that returns a cited report or refuses. Hybrid retrieval over IPCC AR6, indexed on the official MCP Registry | 0 false answers on a frozen 105-question held-out set |
| Incident Evidence Compiler | LLM proposes root causes, a deterministic verifier decides against a content-addressed evidence ledger | RCAEval top-1 0.767 on 90 sealed held-out cases |
| Earthgym | RL environments and evals for weather and climate whose rewards a program can check. Each grader reproduces a published number before it pays a reward | Found 43 ClimaQA (ICLR 2025) test questions word for word in the set the paper fine-tunes on |
| Weather Transformer | 4.8M-parameter ViT for weather forecasting, attention written from scratch in PyTorch | Wind error 22% (u10) and 35% (v10) below persistence on the 2020 test year; temperature still trails anomaly persistence. 129 tests |
| Complaint analytics | Embedding, clustering and hybrid retrieval compared on CFPB data | 200K complaints |
35 merged upstream PRs: Open Climate Fix graph_weather 19 · PyTorch 5 · ExecuTorch 4 · krkn-chaos 4 · MalariaGEN 2 · Brightband ExtremeWeatherBench 1. PyTorch Docathon 2026 Honorable Mention (First-timer), 1 of 7 from 30+ active participants.
Freelance: benchmark tasks for AI coding agents at Shipd by Datacurve. Open to a 6-month internship from December 2026 and full-time roles from July 2027.



