Research

date

slug

research

status

Published

Python, Inspect AI

Scheming detection evaluation testing whether language models exhibit differential transparency based on perceived oversight level.

Built with UK AISI’s Inspect framework. 12 scenarios across 4 categories with model-graded scoring and statistical analysis.

Python, CrossHair, Z3, Hypothesis

Finds bugs in LLM generated code that test suites miss. Tested against benchmark solutions from HumanEval, MBPP, and SWE-Bench.

Uses symbolic execution, property-based testing, and boundary analysis to measure the correctness gap - percentage of passing solutions with hidden bugs.

Python, Docker, Claude/OpenAI, DeepSeek

Multi language formal specification inference benchmark evaluating how well LLMs generate pre/post conditions and invariants.

Supports Java (JML), C (ACSL), Rust (Prusti), and Solidity (SMTChecker) with Docker-based verification. Extended from FormalBench, and added new metrics: minimality, strength.

C, LLVM, KLEE

Research intern, IIT Delhi

Dynamic symbolic analysis tool for Ethereum smart contracts built on KLEE. Detects integer overflow, unchecked send, and contract vulnerabilities through formal verification.

C++

Research assistant, Cryptography and Information Security lab, IISc

Modified MOTION2NX (Mixed-Protocol Secure Two-Party Computation framework) to support multiple parties data computation. Modified Yao’s garbled circuits protocol for three-party computation.

C++, TestU01

Summer intern, NIT Trichy and IEST Shibpur

Analyzed the properties of a cellular automaton (CA) to develop an efficient random Number Generator.

Devised algorithm to generate rules of a 10-state CA PRNG satisfying randomness properties. Verified randomness using Dieharder battery, NIST Statistical Test suite via TestU01: https://github.com/darshana-v/TestU01