Lecturer, Computer Science and Engineering · BRAC University

AI for SWE · LLM post-training · AI safety · Agent orchestration

I did my undergrad in CSE at BUET and now lecture in the same field at BRAC University. My research interest lies at the intersection of four connected areas: AI for software engineering and SRE, LLM post-training, AI safety, and agent orchestration. Before moving into research, I spent several years as a full-stack and AI engineer, shipping production systems.

Based in Dhaka, open to remote research collaboration.

MemAgent: A Cache-Inspired Framework for Augmenting Conversational Web Agents with Task-Specific Information
N. Sakib, P. Barai, S. I. Parisa, A. Iqbal · WOA 2025, From Objects to Agents, Trento, Italy
CEUR-WS Vol-4028, Paper 8

• SREGym #790 · Upstream contribution to the incident-resolution benchmark. A Kafka poison-pill head-of-line block, with an oracle that proves recovery in Kafka itself.
• SREMut · Follow-up to the SREGym work. Mutation testing a benchmark's own verifier to measure what a passing verdict certifies. In progress.

• dc_ops_environment · Physics-grounded datacenter RL environment on OpenEnv: RC thermal networks, continuous multi-objective rewards.
• dc_ops_training · Teacher-distilled SFT to GRPO on AMD MI300X. A multi-signal reward drove +188% composite reward and a 10x per-step gain on hard scenarios.
• GeoQL-4B · Text-to-OverpassQL via SFT then GRPO, with a documented negative result on reward redundancy.
• medical-cot-assistant · Clinical chain-of-thought fine-tuning of a 20B model with QAT, exported to INT4 GGUF.

• AI-safety-forensics · Preregistered forensics separating information from attention in mid-task corrections. Corrective wording cut unsafe deletions to 0/48; naming the file alone did not.
• AI-Security-Testing · Red and blue agents for the Lambda and Berkeley AgentBeats Security Arena. PAIR-style attack refinement, adaptive prompt injection, a call-budgeted defender over A2A.

• scout · Autonomous competitive intelligence. Scheduled agents resolve competitor changes into a bitemporal knowledge graph, then compile cited weekly briefs.
• civilmate-backend · Agentic GraphRAG over building codes. Plans multi-step lookups across a code graph, then checks drawings against site imagery.
• InsightAI-python-backend · Microservices AI platform. Tool-routed generation from PDFs and video, plus CLIP-based multimodal product search.
• Tokkhok-Backend · Personalised RAG for romanised Bengali with a custom transliteration pipeline.

• bd-prescription-medicine-recognize · ResNet50-CRNN for handwritten medicine names. Hash-grouped folds and an SWA ensemble reached 92.53% accuracy, 0.9234 macro-F1.

On Hugging Face @Melikshah.

• GPT-OSS-20B-Clinical-CoT · 4-bit quantised 20B model for clinical chain-of-thought reasoning
• dc_ops_grpo_lora · GRPO-trained LoRA adapter for the datacenter operations environment
• qwen3.5-4b-base-blindspots · Adversarial evaluation set probing Qwen3.5-4B-Base failure modes
• Shajgoj and General Products · 20,000+ item multimodal datasets for image-text retrieval
• Prothom Alo News · Large-scale Bengali news corpus for low-resource NLP

• BRAC University · Lecturer, CSE · present
• Sluqe · Software Engineer, Full Stack and AI · 2025 to 2026
• Oleyn · Founding Engineer, Backend and AI · 2024
• Priyo · Software Engineer, Backend · 2024
• Yobo · Software Engineer Intern · 2024

Languages

Training and inference

Agents and backend

Data and infrastructure

_{Projects · Models · Get in touch}