Research Experience

Agentic Site Reliability Engineering: AI for SRE

2026 | Research Intern, advised by Prof. Tianyin Xu (UIUC)

🔬 Contributing to SREGym, a Benchmark for AI SRE Agents

• Working on the SREGym v2 paper: extending SREGym, an open-source benchmark for evaluating AI agents on cloud-native system failures, from synthetic faults to real-world failure scenarios, and analyzing how agents behave on them.
• Conducting case studies and postmortem analysis of real-world outages, projecting them into reproducible benchmark problems.
• Designing fault scenarios and evaluation oracles that expose documented agent weaknesses: cross-layer reasoning gaps, greedy diagnosis anchoring, and reward hacking.
• Contributed a merged fault scenario reproducing a real Kubernetes admission-control incident, with a reward-hack-resistant evaluation oracle (details on the Projects page).

Code Generation in Bangla: Low-Resource Language Adaptation

2025 | Shared Task at the BLP Workshop, co-located with IJCNLP-AACL

🏆 4th Place out of 32 Teams (submitted system: Pass@1 0.85)

• Fine-tuned Qwen3-4B with parameter-efficient LoRA adapters, chosen over Llama-3.2-3B, TigerLLM, and Qwen2.5-Coder baselines, to overcome data scarcity in Bangla.
• Ablated each intervention on the hidden test set: data augmentation, Chain-of-Thought hints, then step-by-step CoT moved Pass@1 from 0.51 to 0.59. Translating the dataset to English degraded it to 0.50, so the pipeline stayed Bangla-centric.
• Built the execution-guided self-refinement loop that feeds failing code and its execution error back for repair. It lifted the best fine-tuned system to Pass@1 0.62 and, applied to few-shot prompted Gemini 2.5 Pro, produced the submitted result of 0.85, placing 4th of 32 teams.
• Categorized failures by type (ambiguous instructions, constraint violations, test-passing but incorrect solutions) to guide the refinement loop.
• Publication: AdversaryAI at BLP-2025 Task 2: A Think, Refine, and Generate (TriGen) System with LoRA and Self-Refinement for Code Generation

BRAINTEASER: Advanced Commonsense Reasoning in Language Models

2024 | Shared Task at SemEval 2024, co-located with NAACL

• Designed data augmentation pipelines to improve model robustness on commonsense reasoning tasks built around lateral thinking puzzles.
• Conducted a comparative analysis of language models, analyzing performance gaps and reasoning patterns in non-standard logical scenarios.
• Publication: Deja Vu at SemEval 2024 Task 9: A Comparative Study of Advanced Language Models for Commonsense Reasoning

Violence Inciting Text Detection (VITD) in Bangla

2023 | Shared Task at the BLP Workshop, co-located with EMNLP

📊 Ranked 12th in Post-evaluation (19th Official)

• Applied semi-supervised self-training to address class imbalance and improve performance on minority classes.
• Increased dataset diversity through back-translation with the Googletrans API, improving semantic variety while correcting linguistic inconsistencies.
• Implemented an ensemble approach combining multiple transformer models with bagging and majority voting to reduce prediction variance.
• Publication: Team_Syrax at BLP-2023 Task 1: Data Augmentation and Ensemble Based Approach for Violence Inciting Text Detection in Bangla

Improving Answer Space Diversity in Visual Question Answering (VQA)

2022 | Undergraduate Thesis Project

• Conducted a comparative study of VQA methods, identifying core limitations in answer distribution.
• Addressed the “Answer Space Diversity” limitation by augmenting training data with automatically generated, contextually relevant QA pairs using template-based synthesis.
• Demonstrated improved performance on long-tail answer categories, reducing model bias toward frequent answers.