Software developer and researcher · FBK, Trento
I build and investigate software engineering tools, with a focus on AI-assisted testing, software quality and the practical use of large language models.
@article{khandaker2026augmentest,
title = {AugmenTest: A tool for improving test quality through automatic assertion generation},
author = {Shaker Mahmud Khandaker and Fitsum Kifetew and Davide Prandi and Angelo Susi},
journal = {Science of Computer Programming},
year = {2026},
volume = {254},
pages = {103531},
publisher = {Elsevier},
doi = {10.1016/j.scico.2026.103531}
}Generating correct test oracles is still an open problem in automated test generation. AugmenTest asks LLMs to infer the intended behaviour of a method from its documentation and developer comments, without looking at the code, and compares four prompting and retrieval-augmented variants. On 203 bug-revealing tests from 142 Java classes, the Extended Prompt variant produced correct assertions in 30% of cases in the most conservative setting, against 8.2% for the state-of-the-art TOGA approach; the RAG variants did not improve results.
@inproceedings{khandaker2025augmentest,
title = {AugmenTest: Enhancing Tests with LLM-Driven Oracles},
author = {Shaker Mahmud Khandaker and Fitsum Kifetew and Davide Prandi and Angelo Susi},
booktitle = {2025 IEEE Conference on Software Testing, Verification and Validation (ICST), Naples, Italy},
year = {2025},
pages = {279--289},
publisher = {IEEE},
doi = {10.1109/ICST62969.2025.10988926}
}Research
Automatically generated tests often encode what the code does rather than what it should do. This work asks large language models to infer intended behaviour from documentation and turn it into assertions, and measures when that actually helps.
Deriving tests from models of system behaviour, applied to railway interlocking systems.
Using feature models to explore and interactively optimise the configuration space of complex systems.
Selected projects
Research tool · 2024 – 2026
Generates test assertions with large language models from a method's documentation and comments, without reading its implementation. Evaluated at ICST 2025 against TOGA and released as a tool in Science of Computer Programming.
Public AI Challenge 2022 · winning team · 2022
An AI application that classifies the images by theme and detects damage on the photographic plates, built for the Province of Trento's cultural-heritage unit.
Career
Moments
