Naimur
Rahman
Experience
- Developed specialized voice agents supporting both WebRTC and telephony protocols that perform specific workflows like streamlining customer operations, automating FAQ resolution, managing order lifecycle, assisting in shopping, scheduling meetings, and so on. Also integrated a sentiment-aware escalation path that transitions calls to live agents via Amazon Connect when complex needs or negative sentiments are detected.
- Automated scheduler agents task execution by implementing recurrence rules both in the backend and frontend (task scheduling UI).
- Refined prompt instructions both in the system layer and validation layer for the LLM to generate correct T-SQL queries and analyze responses, maintaining schema awareness.
- Developed some agentic workflows and database schemas, reducing fetching latency and ensuring relational integrity.
- Integrated third-party services (Salesforce, New Relic) for data fetching and aggregation as well as implementing webhooks to streamline data synchronization.
- Enabled API caching and worked on WebSocket for real-time responses.
- Developed few frontend components and implemented role-based access control for secure user access.
- Implemented CRUD functionalities in a CMS, ensuring data handling and efficient database management.
- Integrated APIs and improved backend workflows.
- Successfully completed some front-end design tasks for clients.
- Assisted around 40 merchants in payment gateway API integration (Node.js) and resolved their queries.
- Conducted and monitored around 130 User Assurance Testings.
- Managed and maintained merchant records throughout many stages following onboarding.
- Tested and provided reports of Sandbox Environment.
- Conducted in-person and online consultation sessions.
- Assisted students during lab hours alongside the lab instructor.
- Evaluated lab assignments and provided constructive feedback.
Education
Bachelor of Science in Computer Science and Engineering
BRAC UniversityHigher Secondary School Certificate
Govt. Hazi Muhammad Mohsin College
Secondary School Certificate
Chattogram Cantonment Public College
Skills
Programming Languages
Frameworks
Libraries
Runtime Environment
Database
Environment Management
Testing Tools
Version Control
Web Development Concepts
OS
Projects
Full-stack e-commerce platform with product, brand & category management, cart & checkout flow, JWT authentication, admin dashboard with orders/reviews/users management, Cloudinary for image uploads, analytics with Recharts, and PostgreSQL via NeonDB.
Document ingestion pipeline with text chunking & embeddings, Gemini gemini-embedding-001 for vector embeddings, vector similarity search via Upstash, streaming LLM responses, chat history with localStorage persistence, and archive & restore conversations.
Real-time hardware metrics (CPU, memory, disk usage, network speed, and temperature) are fetched from Linux kernel virtual filesystems and displayed in the top panel and dropdown dashboard card. Implemented fully asynchronous I/O and metadata caching to prevent blocking the GNOME Shell compositor thread, and built a customizable settings panel for configuring refresh intervals, units, and layout positions.
Integrated with the D-Bus MPRIS2 interface to track active media players (Spotify, Firefox, VLC) and display real-time track metadata. Built a panel indicator and an iOS-style now-playing dropdown card with scrolling text animations, transport controls, and dynamic seek bars. Implemented asynchronous image downloading and local disk caching for album art to optimize shell responsiveness, and designed a customizable settings panel for positioning and UI customization.
Certifications
Undergraduate Thesis
- Fine-tuned pretrained transformer models (BART, T5, and PEGASUS) on the XL-sum dataset for abstractive text summarization task.
- The fine-tuned PEGASUS model shows a 4.04% improvement in ROUGE-1, a 15.25% increase in ROUGE-2, and a 3.39% rise in ROUGE-L scores, achieving state-of-the-art performance.
- Used SHAP interpreter library for transparency and to help understand the predictions made by the model. The model is deployed on Hugging Face and has been downloaded over 2.8k times.
Honors and Awards
Let's Connect
Always interested in new opportunities, collaborations, and conversations about technology, AI, and software development.