I’m a software engineer at Dynamic Solution Innovators in Dhaka and a research intern in Prof. Tianyin Xu’s group at UIUC. Most of my work is on AI agents for site reliability engineering and software engineering. I also work on NLP for low-resource languages, especially Bangla.

At UIUC, I work on SREGym, an open-source benchmark for testing whether AI agents can diagnose and fix failures in running systems. I’m currently working on its second version, which turns real outages into reproducible scenarios. A large part of my work is building the oracles that decide whether an agent has actually fixed a failure, since an agent can make an error disappear without addressing its cause. One scenario I contributed reproduces a Kubernetes incident where deleting the webhooks clears the error but leaves the underlying problem unresolved.

Outside SREGym, I built an LLM-guided fuzzer for a C XML parser. The model wrote an input generator from a formal grammar. One result I found interesting was that further refinement actually reduced coverage. The feedback loop was rewarding the wrong signals, so making the generator better by those signals did not make it better at reaching new code.

On the NLP side, I’m first author on two shared-task papers at the BLP workshop, including a code generation system for Bangla prompts that placed 4th out of 32 teams at IJCNLP-AACL 2025. I also co-authored a SemEval paper on commonsense reasoning.

My engineering work at DSI often gives me ideas for research questions. I build government systems used across Bangladesh, and I currently work on the data pipeline behind the national voter registration platform, which handles records for nearly 150 million voters. In one incident, a Kafka pipeline had fallen five million events behind. The consumers looked like the obvious place to investigate, but I traced the backlog to slow writes in the storage layer instead.

I also enjoy working close to the JVM. I built and open-sourced gf, a tool that hot-swaps changed bytecode into a running Java server. It cut our edit-deploy cycle from about two minutes to five seconds. I’ve also contributed to PrimeFaces and SREGym, and I teach an online course on enterprise web development with Jakarta EE.

Alongside DSI, I’m a part-time full-stack AI engineer at Altri.ai, where I build RevestAI, a system that scores real estate listings for investors. I split its scoring pipeline into three stages: an LLM extracts features from listing text, an ML model estimates property value, and financial rules handle the calculations. I also rewrote the LLM enrichment step to process listings in bounded parallel batches, making it about ten times faster.

🎓 Education

B.Sc. in Computer Science & Engineering, Shahjalal University of Science and Technology (SUST)
CGPA: 3.53/4.00 (Last two years: 3.83)

🔬 Research & Engineering Interests

• AI for SRE & Software Engineering: benchmarks and evaluation for AI agents that operate real systems, fault injection, and agent behavior analysis
• AI Security & Privacy: security and privacy of AI systems, and AI for vulnerability discovery
• Low-Resource NLP: code generation and text classification for Bangla and other low-resource languages
• AI Infrastructure & Retrieval Systems: LLM enrichment pipelines, RAG systems, and production ML workflows
• Backend & Distributed Systems: enterprise application architecture, workflow engines, observability, and operational tooling

Beyond Work

🏆 Sports

I play football in my free time to stay active and competitive.

📸 Photography

I’m passionate about nature photography: landscapes and everyday moments.

📷 View My Photos