Tanzim Hossain Romel
M.Sc. student, Computing Science · University of Alberta
Currently
- Steering long-horizon coding agents: reproducing TACT, which uses activation steering to curb overthinking and overacting, on SWE-bench Verified.
- BLP 2027 shared tasks (NAACL 2027): adversarial Bangla prompt classification.
- Formal-verification-guided schema synthesis with LLMs: turning English database specifications into relational schemas with verified design contracts.
- Runtime verification using formal verification.
- Human-AI interaction: when AI assistants apply stored user memories, and when they apply them where they do not belong.
I study how AI agents behave in real software systems, and how to make their decisions inspectable, reliable, and trustworthy.
I bring about three years of professional software-engineering experience, formerly at IQVIA.
My research interests are AI4SE, LLM4Coding, trustworthy AI, long-horizon coding agents, and AI for SRE.
I care about focused, sustained work. Here is how I think about work ethic.
I live in Edmonton, Alberta. Fun fact: Edmonton’s North Saskatchewan River Valley is one of the largest stretches of connected urban parkland in North America.
Research
I work on three connected questions.
- Long-horizon coding agents. How can they preserve context, repository memory, and architectural intent as software evolves? See ContextLedger.
- AI for SRE. How should agents be evaluated when diagnosis and recovery happen inside live reliability conditions? See SREGym.
- Trustworthy AI. How can we inspect agent choices, security boundaries, and evidence when the final answer still looks acceptable? See SHIFT.
News
View all- blog
SREGym blog: how we turned real-world postmortems into reproducible SRE benchmark problems. Blog · Tianyin’s post
- education
I have started graduate study in Computing Science at the University of Alberta, in the U-A-Goose group.
- paper
Submitted ‘The Choice Can Be the Attack: Auditing Aligned Backdoors in LLM Agents’ to the Transactions of the Association for Computational Linguistics (TACL) 2026 with Chowdhury Rakin Haider.
- role
Started a remote summer research internship at the University of Illinois Urbana-Champaign through UIUC++ Summer Research in Software Engineering (SRSE) 2026 with Professor Tianyin Xu.
- paper
Resubmitted ‘An Empirical Study on Remote Code Execution in Machine Learning Model Hosting Ecosystems’ to ACM Transactions on Software Engineering and Methodology (TOSEM) 2026. Preprint available on arXiv.
Latest posts
View allSelected publications
View all- An Empirical Study on Remote Code Execution in Machine Learning Model Hosting Ecosystems
Preprint available on arXiv
A cross-platform study of model-loading risk across five ML hosting ecosystems, using static analysis, malware-signature scanning, and analysis of more than 600 developer discussions.
- The Choice Can Be the Attack: Auditing Aligned Backdoors in LLM Agents
Submitted to TACL 2026; manuscript not public
SHIFT, a known-trigger audit for backdoors that make an agent prefer one valid option over another while the final answer still looks acceptable.
- Sentiment Analysis of Anonymous Crisis Reports in Bangladesh
Submitted to the ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW)
A multilingual analysis of anonymous, crowd-sourced reports from Bangladesh's 2024 national crisis.
Selected work
View all- ctxhelm: agent-native context compiler
Local-first context plans, release proof, and source-free agent evaluation
thromel/ctxhelm - PatchSmith: a harness for repair agents
System design and R&D notes for bounded patches, sandbox validation, and auditable repair runs
thromel/patchsmith - ContextLedger: context accounting for coding agents
A research notebook on OpenCode compaction, provenance, and live GPT-5.5 experiments