Ahmed Anwar

• E-KYC Plug-and-Play Solution (ekyc.nagorikcard.com): Developed a comprehensive E-KYC platform for institutional use. Engineered and integrated modules for robust NID OCR, automated NID verification, face liveness detection, and highly accurate face matching to streamline user onboarding.
• Pensioner Verification System (pension.nagorikcard.com): Built an automated verification system utilizing face liveness detection and matching to periodically verify the life status of TCB pensioners, ensuring secure and accurate pension distribution.
• ID Document Liveness Detection System: Developed a custom vision model capable of classifying multiple spoof types including genuine, photocopy, screen replay, reprint/critical spoof, half card, and blurry images of TCB Smart Family Cards issued by Government. This enhanced security measures for ID verification workflows.
• High-Performance ML Inference: Optimized Vision Transformer models for production, achieving sub-150ms latency on CPU-only environments. Scaled system throughput to 20+ requests per second (RPS) by implementing multi-threading, Nginx load balancing, and efficient model selection.
• Bengali News Insight Engine: Designed a full-scale data pipeline using Apache Airflow to scrap 40+ Bengali newspapers daily. Applied NLP and vector search techniques to extract and retrieve relevant insights using semantic similarity search.
• Bengali STT & TTS Models: Achieved notable progress in building a custom speech-to-text (STT) model for the Bengali language, enhancing language accessibility and enabling voice-powered interfaces; currently refining STT performance before initiating the text-to-speech (TTS) phase of the project.
• Agentic Automation Systems: Developed AI agent-based systems capable of automating workflows such as chat-driven information retrieval, report generation, and dynamic system updates—reducing manual overhead and improving operational efficiency.
• Model Fine-Tuning & Optimization: Fine-tuned large language models using Parameter-Efficient Fine-Tuning (PEFT) strategies to adapt them to specialized business domains with minimal resource consumption.
• AI Infrastructure on AWS: Managed full ML model lifecycle using AWS SageMaker and Bedrock for training, deployment, and auto-scaling across production workloads. Ensured system robustness and reliability through cloud-native monitoring and optimization.
• Semantic Vector Search Systems: Designed and deployed vector search pipelines using state-of-the-art embedding models and vector databases to enable semantic retrieval, chatbot memory, and content-based recommendation across multilingual data.
• Backend Engineering & DevOps: Built scalable backends primarily with Django and occasionally FastAPI. Set up Nginx for load balancing, implemented CI/CD pipelines, containerized applications with Docker, and monitored live deployments to ensure system reliability and performance.

Promoted from Python Intern to Junior Programmer based on performance and key contributions to AI agent development and workflow automation systems.

• Designed and refined advanced AI Agent architectures for task automation and real-time decision-making. Built the OwnAgent system that learns and replicates user workflows using LLMs and automation primitives.
• Worked on Retrieval-Augmented Generation (RAG) pipelines using LLMs for domain-specific information retrieval. Delivered an AI system trained with Llama2 to answer queries via context-based retrieval.
• Built and deployed practical Python solutions using FastAPI. Developed a LinkedIn Automation System to automate login, engagement, and content analysis workflows.
• Collaborated in code reviews and team engineering sprints, maintaining high code quality and fast iteration cycles.

• Tutored Python, OOP, Data Structures, and algorithms at BRAC University.
• Provided individualized guidance, fostering a supportive learning environment.
• Developed time management, teaching and leadership skills.

• Designed a classification model to differentiate between Rabindranath Tagore’s texts and non-Tagore texts.
• Analyzed stylistic and linguistic differences between Rabindranath-era texts and contemporary Bangla texts.
• Conducted multi-author text classification for Bangla literature.
• Fine-tuned mT5 and BanglaT5 models to generate texts in Rabindranath Tagore’s writing style.
• Developed a classifier to distinguish between authentic Tagore texts and AI-generated imitations.