Agent governance
Agent GovBench
94.7% vs 29.8%On 999 synthetic scenarios, Gemini Flash blocked 94.7% of misuse and flagged 0.4% of benign calls. Model Armor at MEDIUM blocked 29.8% and flagged 27.1%.
Agent Gateway was not tested.
The 60-second path
AI Engineer · Agentic AI and Evaluation
Brejesh Balakrishnan
AI Engineer at Cloud202, since Sep 2026
I build governance and safety controls for AI agents on Amazon Bedrock and AgentCore. I publish my evaluations, including the misses.
Three cases
Bounded RAG
Compliance Workbench
13 of 14Across 3,946 provisions from nine official texts, the assistant declined 13 of 14 out-of-scope questions without a model call and wrongly declined none of 50 covered questions.
One out-of-scope question got an answer. No legal advice.
Multilingual routing
Mozhi
0.941 at 2.2 msA 23 MB router scored 0.941 macro-F1 (95% CI 0.916 to 0.962, n = 2,805), versus 0.490 at 3,333 ms for the original prompt. False actions fell from 40.7% to 0.8%.
Tamil and Malayalam missed the 0.85 target at 0.818 and 0.809.
Experience
- Sep 2026 to Present
Cloud202 Limited · AI Engineer. Audited 15+ agent governance features against AWS sample projects.
- Jul 2024 to Mar 2026
Accenture · Advanced Application Engineering Analyst. Cut slow API responses from roughly 80% to under 20%.
- 2021 to 2023
Four internships in NL-to-SQL, machine learning, research proposals, and sponsorship.
- 2020 to 2024
B.E. Electronics and Communication Engineering, College of Engineering Guindy, Anna University. CGPA 8.54.
How I work
Measure the baseline, set the pass mark before scoring, check independently, and publish the misses.