Independent researcher in agent evaluation and agentic systems.

B.Tech, Computer Science and Engineering, National Institute of Technology Patna. Pune, India.

Research focus: evaluation integrity for agent benchmarks. Outcome tests alone can certify no-op methods, cosmetic adaptation, and fixture edits. Technical report When the delta is fake proposes a pass-bar for task authorship and a construction that catches fake deltas under ranking AUROC (mean-shift invariance + train-site shortcut flip + frozen verifier-recomputed control). DOI: 10.5281/zenodo.23829461.

Also works on eval inside product agent loops (draft → score → revise) and reliability through tool layers (cancel, timeouts, evidence grounding).

• agent-eval-suite: reference tasks and the technical report on fake deltas in agent evals. One-command repro.
• workspace-brief: brief a live GitHub workspace with a planner, citations, memory and JSON traces.
• brief-desk: CopilotKit + AG-UI chat desk with MCP tools for workspace and corpus briefing.

Selected merged fixes:

• celery: don't fail a task whose SUCCESS was stored before the worker was lost, don't record FAILURE when on_success raises after SUCCESS
• kombu: stop channel retries at the connection max_retries bound
• asyncpg: bound close(timeout) around a pending cancel
• go-yaml: end tags at flow indicators in flow content
• mlx-lm: fix mlx_lm.server model-swap leak and dead generation thread
• gunicorn: stop writing a 500 on top of a response that already started

In review on eval harnesses: Harbor #4003, Terminal-Bench #2480. All merged PRs.