B.Tech AI/ML Student · Preparing for AI Safety Evaluator / Agent Safety & Safeguards roles

I am building the practical skills needed to evaluate tool-using AI agents safely and reproducibly. My focus is turning safety policies and tool-use risks into clear test cases, metrics, graders, mitigations, and regression suites that another person can independently run and verify.

I design controlled evaluations, measure both safety and useful-task performance, and document limitations honestly.

• Translating safety policies into measurable behaviors
• Designing realistic test cases (benign, harmful, ambiguous, over-refusal)
• Running controlled model and agent evaluations with full experimental metadata
• Building deterministic and model-based graders
• Calculating core safety and usefulness metrics (attack success rate, correct-refusal, over-refusal, task completion)
• Proposing and testing mitigations while checking for new trade-offs
• Creating held-out regression tests and one-command reproducible suites
• Writing clear technical reports with honest limitations

Lightweight Python for evaluation work Variables, functions, CSV/JSONL handling, API calls, error handling & retries, saving prompts/responses/parameters, basic pandas, simple pytest assertions

Model & Agent Fundamentals Tokens & context windows · System / developer / user / tool instructions · Refusal vs safe-completion behavior · Direct & indirect prompt injection · Multi-turn manipulation · Tool selection & authorization · Memory poisoning · Sensitive-data disclosure · Under-refusal & over-refusal · Trusted vs untrusted content

Evaluation Design Threat models · Behavioral specifications · Positive & negative controls · Binary / ordinal / multi-label rubrics · Deterministic, model-based, and human graders · Development vs held-out sets · Repeated trials · Full reproducibility

Basic Statistics for Safety Work Success/failure rates · Attack-success rate · Correct-refusal & over-refusal rates · Task-completion rate · Precision/recall · False positives/negatives · Confidence intervals

I previously built practical Android applications (Jetpack Compose, foreground services, overlays, state machines, privacy controls, offline fallbacks, unit-tested behavior engines).

That work trained me to:

• Prefer small, understandable systems with explicit boundaries
• Document what a system can and cannot do
• Turn vague requirements into testable behavior
• Write tests and keep clear records of limitations

These habits map directly to evaluation design and reproducible safety work.

(Evaluation harnesses and safety test suites are in active preparation and will be added here as they become reproducible.)

• Clear architecture and documented limitations
• Unit testing and modular design
• Full recording of prompts, responses, parameters, and timestamps
• Preference for work another person can independently reproduce
• Honest reporting of remaining risks and over-refusal trade-offs

• Controlled evaluation pipelines for tool-using agents
• Grader construction and validation
• Mitigation experiments with safety–usefulness measurement
• Regression suites that run with one command
• Professional evaluation reports with threat models, methodology, results, and limitations

• GitHub: @maxew6
• X: @maheshngxt
• Email: maheshkonda564@gmail.com
• Location: India

This profile is a record of deliberate skill-building toward AI Safety Evaluator work — focused on reproducibility, clear measurement, and honest documentation of limitations.