AI evaluation and benchmark engineering · Agentic AI security · Python
Bengaluru, India · LinkedIn · Open-source contributions
I investigate why agent workflows fail and build the tests that make those failures reproducible. My work spans terminal benchmarking, reference-solution and verifier validation, Docker environments, and security evaluation of agent tool pipelines.
I hold an M.Tech in Cybersecurity and a B.E. in Computer Science and Engineering from Visvesvaraya Technological University. My current contract work includes benchmarking on the Snorkel AI platform via AirDawg Labs, long-horizon projects at AfterQuery, and AI evaluation at DataAnnotation.
- Evaluation reliability: inspect execution traces, test results, and final artifacts; distinguish model failures from specification, reference, verifier, or environment defects.
- Reproducible debugging: isolate dependency and runtime problems with Python, Shell, Docker, and targeted regression tests.
- Agentic AI safety: study MCP tool boundaries, prompt injection, runtime guardrails, and the integrity of evaluation inputs.
| Project | Focus | Where to start |
|---|---|---|
| Agentic security research | Public demonstration artifacts from my M.Tech work on MCP and agent security. | Research context, script behavior, and experiment boundaries. |
| Payroll anomaly detection | Experimental Python pipelines for anomaly scoring, evaluation, and drift analysis. | Batch and streaming workflows, evaluation implementation, and limitations. |
| VulnHawk | Python security-scanning project with an asynchronous CLI and JSON reporting. | Implemented scan modules, local setup, and architecture. |
ScanCode Toolkit #5292 — dependency regression
Traced a Click 8.5.0 default-value change that silently enabled an output filter. Added an explicit option default to preserve function-API behavior.
ScanCode Toolkit #5279 — CLI compatibility
Resolved a conflicting short option while preserving the established copyright-scan flag, and updated the CLI help fixtures and documentation.
Both changes were merged upstream. My pull-request history also includes ongoing reviews and proposed fixes across Python tooling and software supply-chain projects.
| Area | Tools and methods |
|---|---|
| Evaluation and testing | Harbor, Terminal-Bench, pytest, unittest, reference and null-agent controls, rubric design |
| Development | Python, Bash/Shell, TypeScript, SQL, Git |
| Environments | Docker, Linux, GitHub Actions, uv, virtual environments |
| Agent workflows | Claude Code, OpenAI Codex, MCP, trajectory analysis, multi-step verification |
| Data analysis | Pandas, NumPy, anomaly evaluation |
Open to remote contract work in AI evaluation, benchmark engineering, and agent security. Connect on LinkedIn.