research
Research on reliable AI agents, SRE-agent evaluation, ML ecosystem security, agent-choice auditing, and developer tools.
Research dossier
AI agents in real software systems
I study AI agents where software systems, security boundaries, and operational evidence matter. The aim is to make agent behavior inspectable before it is trusted. Current work combines reliability-agent evaluation, security audits for agent choices, and developer tools that make context and software evolution easier to examine.
Publication record CV Scholar profile
Questions, artifacts, and what the evidence supports
Evidence anchors
01 · Reliability-agent evaluation
How can we evaluate an SRE agent in a live software environment without mistaking a shallow recovery for a reliable one?
SREGym: AI for SRE and reliability-agent evaluation
Through UIUC++ SRSE 2026, I am working on SREGym, a benchmark for evaluating SRE agents in live system environments and real-world reliability problems.
| Status | Active research · UIUC++ SRSE 2026 |
|---|---|
| Context | June–August 2026 · Professor Tianyin Xu and the UIUC systems reliability group |
| Last verified |
02 · ML ecosystem security
What risks emerge when model-hosting platforms load community code and artifacts at ecosystem scale?
Remote Code Execution in ML model hosting ecosystems
A cross-platform empirical study examines custom model-loading risk across five ML hosting ecosystems, using static analysis, malware-signature scanning, and more than 600 developer discussions.
| Status | Under review at ICSE 2027 |
|---|---|
| Context | With Mohammed Latif Siddiq, Natalie Sekerak, Beatrice Casey, and Joanna C. S. Santos |
| Last verified |
03 · Agent security
Can a trigger steer an agent toward an attacker’s preferred valid choice while its final answer still appears acceptable?
The Choice Can Be the Attack: auditing aligned backdoors in LLM agents
SHIFT compares matched tasks with and without a known trigger, then asks whether a changed valid choice favors an attacker’s target after accounting for ordinary option quality.
| Status | Submitted to TACL 2026 · manuscript not public |
|---|---|
| Context | With Chowdhury Rakin Haider · abstract-level summary only |
| Last verified |
The manuscript is not public; this record is deliberately limited to an abstract-level summary.
04 · Coding-agent systems
How can coding agents preserve exact session facts, inspect the right files, and leave testable repair evidence?
Context, retrieval, and auditable repair
ContextLedger measures what survives context compaction, ctxhelm and HelmBench evaluate retrieval efficiency, and PatchSmith records bounded repair runs with sandbox validation and inspectable artifacts.
| Status | Active systems research |
|---|---|
| Context | ContextLedger · ctxhelm and HelmBench · PatchSmith |
| Last verified |
05 · Agent memory
Can coding agents preserve architectural intent while developers work through fast natural-language programming loops?
History-aware vibe coding
This emerging direction studies how agents should use commit history, prior edits, repository memory, and design intent without losing the speed of vibe coding.
| Status | Emerging research direction |
|---|---|
| Context | With Dr. Zhou Yang · University of Alberta and Amii |
| Last verified |
The instrumentation around the research
Supporting systems
These systems make the agent’s context, choices, and program-level effects more inspectable.
| System | Status | Role |
|---|---|---|
| ContextLedger | Research system | Typed, provenance-carrying session facts for measuring context compaction. |
| ctxhelm and HelmBench | Open-source tooling | Guidance and evaluation for whether coding agents inspect the right files with less irrelevant reading. |
| RefactoringMiner MCP | Open-source collaboration | Program-analysis access that lets coding agents inspect refactorings and AST diffs beyond raw patches. |
Work with me on an artifact, question, or evaluation
Collaboration
If you are studying dependable AI agents or building reliable software systems, I would be glad to compare research questions, evaluation evidence, or engineering tradeoffs.