Nazmus Sakib

AI Engineer & Researcher

RL post-training · Agentic systems · Full-stack AI products

I build language-model systems end to end - the reinforcement learning that trains a policy, the agent architecture that puts it to work, and the product it ships inside.

My research is in RL post-training: reward design, GRPO and its variants, and the custom environments that make verifiable RL possible in domains where no benchmark exists yet. I'm drawn to the part usually skipped - building the simulator, the verifier and the reward signal, then measuring honestly whether RL earned its compute over supervised fine-tuning.

The engineering half is agentic. Memory architectures that let agents carry context across sessions (the subject of my WOA 2025 paper), tool calling and multi-step orchestration, graph and hybrid retrieval, and the adversarial testing that shows where those tool chains break. Around all of it sits the ordinary work that decides whether a model ever reaches a user: inference services, vector search, auth, payments, realtime web and mobile clients, and the CI that ships them. I also care about low-resource language evaluation, particularly Bengali.

CSE, BUET. 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, 26th Workshop From Objects to Agents, Trento, Italy. CEUR-WS Vol-4028, Paper 8

An agent memory architecture: a Memory Cache Bank with time-based expiration that decouples information gathering from task execution, so an agent stops re-asking users for details it has already learned. Reduces average conversation turns by 22.4% (5.00 → 3.88) across 150 Mind2Web tasks; a 15-participant study showed a 58% reduction in completion time for recurring tasks.

Published on Hugging Face @Melikshah.

_{Projects · Models · Get in touch}