Applied Research Engineer · Abundant (YC)
Neel Harsola
I do research on agent evaluations at Abundant (YC) — designing benchmarks, eval harnesses, and the infrastructure that measures how frontier AI agents reason, use tools, and write code. My current focus is long-horizon autonomy: whether agents can carry real software projects across hours and days, not minutes.
I'm a co-author of SWE-Marathon, an ultra long-horizon software-engineering benchmark whose results have been cited on the official model cards of Kimi K3, Grok 4.5, and GLM 5.2.
On the model cards
Frontier labs cite SWE-Marathon in their official model cards.
Kimi K3
Moonshot AI
Jul 2026
Grok 4.5
xAI
Jul 2026
GLM 5.2
Zhipu AI
Jun 2026
Work
Benchmark · Abundant AI
SWE-Marathon
Do agents autonomously complete ultra-long-horizon software work?
An ultra long-horizon software-engineering benchmark. Agents are dropped into real multi-language codebases and asked to carry entire projects to completion — with a public leaderboard, 320 GB of fully open trial logs, and reproducible runs on the Harbor framework.
Currently
- Long-horizon coding agents — how far can an agent go unattended?
- Agent benchmarks and eval harnesses for terminal-based agents
- Eval infrastructure powering 5 of the 6 top frontier labs and multiple Fortune 500 companies
Experience
Applied Research Engineer · Abundant (YC)
2025 —
Building the benchmarks and evaluation infrastructure behind SWE-Marathon and related agent evals — designing end-to-end harnesses that test how agents reason, use tools, and execute code inside containerized sandboxes, with benchmark datasets drawn from real-world debugging and engineering scenarios. This work powers evaluations for 5 of the 6 top frontier labs and multiple Fortune 500 companies.
Full Stack Engineer · MSBC Group
2025
Enterprise AI: RAG platforms, a citation-backed legal knowledge assistant, and financial scenario tooling.
Software Engineer · Clarity Sphere Technologies
2023 — 25
ETL pipelines on Glue and Airflow; LLM workflows with LangChain/LangGraph; custom agent frameworks.
Education
B.E. Computer Engineering
L.J University, Ahmedabad
2022 — 25
Diploma, Computer Engineering
A.V. Parekh Technical Institute
2019 — 22
Say hello — work.neelharsola@gmail.com