Ahasan Kabir

Biography

I am a Ph.D. student in Computer Science at the University of Central Florida, advised by Prof. Qian Lou. My research focuses on efficient AI inference — building systems that maximize GPU and hardware utilization for large-scale LLM serving through hardware-aware routing, scheduling, and orchestration.

My first-author paper HW-Router, on hardware-aware routing for scalable multi-LLM serving, appeared at DAC 2026 (Design Automation Conference), and I was selected as a DAC Young Fellow in 2026. Most recently, I built InfraMind, an infrastructure-aware multi-agent orchestration framework.

News

• Aug 2026I am looking for research internships for Summer 2027 in ML systems and efficient inference.
• Jun 2026InfraMind, our infrastructure-aware multi-agent orchestration framework, is now on arXiv.
• 2026Selected as a DAC Young Fellow 2026.
• 2026HW-Router is accepted at DAC 2026, the 63rd ACM/IEEE Design Automation Conference.
• Aug 2025Started my Ph.D. in Computer Science at the University of Central Florida.

Selected Publications

1. Efficiency HW-Router: Hardware-Aware Routing for Scalable Multi-LLM Serving

   Ahasan Kabir, Jiaqi Xue, Mengxin Zheng, Qian Lou

   DAC 2026 · 63rd ACM/IEEE Design Automation Conference PDF

   Existing routers choose among LLMs using static model characteristics, missing the performance swings caused by hardware load and contention. HW-Router brings live GPU signals — queue length, cache usage, recent latency — into model selection, cutting end-to-end latency by 3.4–3.9× and raising SLO attainment by 46–48 pp over state-of-the-art routers, at ~200 µs overhead.

2. Agents InfraMind: Infrastructure-Aware Multi-Agent Orchestration

   Ahasan Kabir, Jiaqi Xue, Mengxin Zheng, Qian Lou

   arXiv 2026 · preprint arXiv

   Multi-agent LLM systems plan and route from task and model features alone, so preferred models build deep queues while equally capable ones sit idle. InfraMind makes planning, model selection, and scheduling aware of live infrastructure state, trained end-to-end with reinforcement learning — up to +7.6 pp higher accuracy with 7× lower latency, and 99.9% SLO compliance under load where every baseline drops below 50%.

3. NLP BEmoLexBERT: A Hybrid Model for Multilabel Textual Emotion Classification in Bangla by Combining Transformers with Lexicon Features

   Ahasan Kabir, Animesh Roy, Zaima Taheri

   BLP @ EMNLP 2023 · Workshop on Bangla Language Processing PDF ACL

4. Multimodal BEmoFusionNet: A Deep Learning Approach for Multimodal Emotion Classification in Bangla Social Media Posts

   Zaima Sartaj Taheri, Animesh Chandra Roy, Ahasan Kabir

   ICCIT 2023 · IEEE PDF DOI

Honors & Awards

• DAC Young Fellow, ACM/IEEE Design Automation Conference — 2026
• ICPC Asia West Continent Finalist — 22nd place (2022) and 25th place (2021), representing CUET
• Codeforces Candidate Master — max rating 2094 (top 1%), 3300+ problems solved
• Meta Hacker Cup 2021 — 38th of 34,584 in the Qualification Round; 116th of 12,692 in Round 1

Experience

• 2025 – now Graduate Research Assistant, University of Central Florida — advisor: Prof. Qian Lou
• 2023 – 2025 Software Engineer, Kinetik Healthcare Solutions — New York, USA (remote). Built the user-facing ride-booking iOS app that patients use to book and track medical transportation, published on the App Store.

Education

• 2025 – now Ph.D. in Computer Science, University of Central Florida
• 2018 – 2023 B.Sc. in Computer Science & Engineering, Chittagong University of Engineering & Technology (CUET)