Skip to content
View MGpromax's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report MGpromax

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
MGpromax/README.md

Manoj Gowda

AI Benchmark & Evaluation Engineer · Bengaluru, India

I build the tasks that break AI agents.

I design terminal-based agentic benchmarks in the Harbor / Terminal-Bench format, calibrated so frontier models fail more often than they pass. B.E in Computer Science and Engineering and M.Tech in Cybersecurity @ VTU.


Benchmark authoring

  • 280+ Terminal-Bench tasks authored, ~270 accepted — Docker-isolated environments, deterministic test oracles, canary-string hygiene
  • Difficulty calibration against frontier models: iterative hardening until pass rates land at ≤3/5 runs (medium/hard thresholds)
  • Anti-pattern-matching design — tasks built so models must actually reason through the environment instead of recognizing a shape they've seen before
  • Milestone and non-milestone task formats, multi-step verification, reproducible harnesses

Evaluation work

  • 500+ completed evaluation tasks across Snorkel AI, Appen, Toloka, Mindrift, AfterQuery, CrowdGen — consistently strong quality ratings
  • Code and PR evaluation workflows, IDE-arena dataset creation, model output rating and ranking
  • Agent-config authoring (CLAUDE.md) for reproducible, repeatable task-creation pipelines

Toolkit

Python Bash TypeScript Docker pytest Linux Git Harbor Terminal-Bench MCP LLM evaluation tmux asciinema


What I care about

  • Benchmarks that resist pattern-matching rather than reward it
  • Deterministic, reproducible eval harnesses — no flaky oracles
  • Honest difficulty calibration: a task is only "hard" if the numbers say so

Popular repositories Loading

  1. Cyber-security---Task-1 Cyber-security---Task-1 Public

    Python

  2. Cyber-security---Task-2 Cyber-security---Task-2 Public

    Python

  3. Cyber-security---Task-3 Cyber-security---Task-3 Public

  4. Cyber-security---Task-4 Cyber-security---Task-4 Public

  5. Cyber-Security---task-5 Cyber-Security---task-5 Public

  6. Cyber-security---Task-6 Cyber-security---Task-6 Public