Ameya Srivastava

This portfolio acts as my public academic CV. You can save it as a PDF (or press Ctrl+P / Cmd+P).

Research Interests. Security evaluation and analysis of complex software and intelligent systems, with an emphasis on methods for systematically discovering and reproducing their anomalous behaviors, failure modes, and vulnerabilities under adversarial conditions.

Keywords: ML for Cybersecurity, Adversarial AI, Malware, Fuzz Testing, Reasoning Models, LLMs

Affiliations: Independent (present), SVNIT Surat

I am currently applying and looking for research-oriented graduate student, research assistant, and internship roles.

Education

S.V. National Institute of Technology (NIT), Surat (India) — (Nov. 2022 - May 2026)
- B.Tech, Computer Science and Engineering (CSE)
- Grade: 8.64/10.00 CGPA, Rank: 20/125 (First Class with Distinction)
- Activities: ACM, Google Developers Student Club

Research and Work Experience

ML Evaluations Contributor (Freelance) — Shipd by Datacurve (YC W24) — (Mar. 2026 – Present)
- Curating ML benchmarks, evaluation datasets, and model-engineering tasks to evaluate reasoning and agentic capabilities in frontier models.
Research Intern — S.V. National Institute of Technology, Surat — (Jan. – May 2026)
- Advised by Prof. Devesh C. Jinwala, designed and implemented a controlled evaluation of representation-architectures for behavioral malware detection (led to the ICDCIT 2027 paper).
Summer Intern (AI) — Sparkflows.io — (May – Jul. 2025)
- Developed an agentic pipeline to detect drug-drug interaction in comorbid patient cases and optimized deployed LLM workflows for reliability and reproducibility.

Selected Projects

Cascaded Intent Development Script (CIDEr) — (Nov. 2025)
- An intentful domain-specific language (DSL) that uses an AI-backed compiler pipeline which translates declarative logic into verified executable Python/C through intermediate representations after deterministic structural linting. This project is at present a working prototype.
- [GitHub]

Competitions, Hackathons and Writeups

AI Agent Security - Multi-Step Tool Attacks (OpenAI, IEEE, Kaggle 2026) — Developed an attack algorithm to exploit weaknesses in an AI agent system simulation defended by a private guardrail. Ranked 536 out of 4,186 teams as a solo participant.
Amazon ML Challenge (2024) — Led a team and built a hybrid NER pipeline over ~132,000 noisy and diverse product image samples. Ranked 155 out of 2,438 teams among 70,000+ participants.
LLM’s – You Can’t Please Them All (Kaggle 2024) — Achieved a score equivalent to ~53% inter-judge disagreement against a black-box 3-judge LLM panel (that evaluated written essays) in an adversarial AI challenge.
Red-Teaming Challenge - OpenAI gpt-oss-20b (Kaggle 2025) — Introduced an adversarial fuzzing framework for language models, and used it to develop inputs to increase harmful compliance with malicious requests in a target OSS reasoning model at non-trivial rates [Writeup].
Measuring Progress Toward AGI - Cognitive Abilities (Kaggle 2026) — Introduced a benchmark to evaluate the epistemic restraint and abstention capabilities of frontier models on synthetically generated clinical null-signal adversarial inputs [Writeup].