Md Sifat Hossain

Dhaka, Bangladesh

Supervisor: Md. Fahim Arefin (DU CSE), in collaboration with Prof. Tarannum Shaila Zaman (UMBC, Information Systems).

• Designed and implemented LLM-ProS, a novel evaluation framework for benchmarking LLM performance on ICPC World Finals problems. Curated a 166-problem dataset (2011-2024), built automated submission pipelines via Codeforces Gym, and analyzed five state-of-the-art models across correctness, resource utilization, and chain-of-thought reasoning. Published at ICSE 2025 (LLM4Code).
• Extended this work into A-ProS, an autonomous multi-model agentic framework that separates solution generation from specialized debugging feedback under a 2×3 factorial design. Developed the full orchestration pipeline, Codeforces browser automation (Selenium + Playwright), verdict capture, and SQLite logging. Ran ablations on persistent vs. stateless context and trust calibration (ECE) across critic models. Accepted at ACM TOSEM 2026.
• Completed undergraduate thesis “A Hybrid LLM Feedback Framework for Automated Competitive Programming Workflows,” proposing an iterative test-driven benchmarking pipeline integrating OpenAI o3-mini with specialist LLMs (DeepSeek, Qwen) for error diagnosis and code refinement via Codeforces-based validation.

Remote

• Contributed to Reinforcement Learning from Human Feedback (RLHF) data creation supporting large-scale AI model alignment research, focusing on dataset quality, consistency, and annotation methodology for code and reasoning tasks.
• Led a team of 10, establishing quality control protocols that ensured annotation reliability across diverse task domains.
• Coordinated cross-functional team activities and maintained consistency standards critical to downstream model training, directly supporting AI alignment objectives.

Remote, India

• Developed a fraud detection model applying anomaly detection and supervised learning on imbalanced datasets, achieving 85% precision. Investigated model behavior under class imbalance and precision-recall trade-offs.
• Automated model training and deployment pipelines using Docker and Jenkins, enabling reproducible ML experimentation and continuous integration of model updates.

Dhaka, Bangladesh

• Develop and maintain scalable features for Therap’s EHR SaaS platform (used across all 50 US states) using Java, Spring, Hibernate, JSP, and Oracle DB, with a focus on correctness and reliability under HIPAA constraints.
• Build and containerize full-stack modules using React.js and Docker, deployed on WebLogic Server, contributing to platform stability and consistent delivery across environments.

Remote, London, UK

• Built and deployed scalable web applications for 5+ clients using React.js and FastAPI, significantly reducing average page load times through targeted performance optimization.
• Architected CI/CD pipelines with automated testing and deployment workflows on AWS (EC2, S3, RDS), accelerating release cycles while maintaining production code quality.

Dhaka, Bangladesh

• Organized the 6th National English Carnival, a national-level academic competition with 10,000+ participants across 30+ events; previously co-ordinated the 5th edition (5,000+ participants).
• Owned judge coordination and external communications, managing scheduling and briefing for judges drawn from industry, academia, and senior education across parallel event tracks.
• Led sponsorship outreach, securing corporate partnerships that funded event operations and prize pools.

Reference implementation of A-ProS (TOSEM 2026) - an autonomous agentic framework separating solution generation (GPT-4 / GPT-5) from specialized debugging feedback (DeepSeek-R1, Llama-3.3, Codestral) under a 2×3 factorial design. Persistent multi-model feedback loops achieve 2.2-2.3× greater gains than stateless baselines on 367 ICPC and Codeforces problems.

Data and orchestration pipeline underlying LLM-ProS (ICSE 2025) and the broader A-ProS benchmark - scraping 166 ICPC World Finals problems, normalizing LaTeX/HTML, structuring statements / I/O specs / constraints / samples, automating Codeforces Gym submissions, and capturing per-attempt verdict + runtime + memory in SQLite. Extended to Codeforces, forming the 367-problem benchmark used in A-ProS.

Scraper extracting video descriptions and author metadata for specified keywords and tags, enabling structured analysis across 5,000+ videos. Selenium for dynamic rendering, BeautifulSoup for parsing, SQLite3 for storage - the same scraping architecture later applied in the LLM benchmark pipeline.

Multi-role event management platform with pessimistic locking to guarantee transactional consistency and prevent concurrent booking conflicts at scale. Real-time event filtering with asynchronous data retrieval for responsive search.

Job search platform with automated CV generation and personalized job matching, integrating live job scraping to fetch and rank relevant listings.

Personality-based social networking Android app with null-safe Kotlin architecture, reducing crash rates and improving runtime stability.