Ekramul Islam
Verified Expert in Engineering
Machine Learning Engineer and Developer
Dhaka, Dhaka Division, Bangladesh
Toptal member since November 19, 2025
Ekramul is an accomplished machine learning (ML) engineer with experience at a Fortune 500 company. He specializes in developing complex AI agents with LangGraph and deploying open-source large language models (LLMs) with low latency. Ekramul's research yields state-of-the-art results in NLP tasks, including shallow parsing, sentiment analysis, named entity recognition, and more, resulting in publications at top-tier conferences such as KDD, NAACL, and EMNLP.
Portfolio
Experience
- Scikit-learn - 5 years
- Python - 5 years
- PyTorch - 4 years
- Hugging Face Transformers - 4 years
- FastAPI - 3 years
- TensorFlow - 3 years
- LangGraph - 1 year
- ChromaDB - 1 year
Work Experience
AI Engineer Python Developer (via Toptal)
P95 SARL
- Built ReAct LLM agents with tool-calling, prompt versioning, and guardrails to classify, enrich, and score events from RSS, REST, and SQL sources, using hybrid rule/LLM scoring with feedback recalibration.
- Developed FastAPI services and PostgreSQL schemas with Alembic migrations, pgvector embeddings for entity resolution, Docker deployment, and scheduled ingest/enrich pipelines with deduplication and retries.
- Delivered a Streamlit dashboard with filterable signal feeds, score breakdowns, up/down feedback capture, source health monitoring, and semantic search over enriched signals.
Machine Learning Engineer
IQVIA
- Designed and built an enterprise agentic AI assistant with multi-agent orchestration (LangGraph) for autonomous task planning, reasoning, and execution, enabling users to ingest, query, and analyze large-scale datasets through natural language.
- Developed specialized agents and ML tools for dynamic clustering, feature selection, text-to-SQL, anomaly detection, and RAG-backed analytics, turning complex big data workflows into conversational insights with tables and visualizations.
- Fine-tuned and deployed foundation models (Llama 2, Llama 3) and open-source LLMs (DeepSeek, Qwen) on AWS SageMaker with vLLM, optimizing inference to 90ms first-token latency for scalable, cost-efficient production use.
- Led end-to-end delivery with clients and cross-functional teams — translating product requirements into production-ready AI on AWS, integrating with enterprise data platforms via secure, modular cloud deployment.
Machine Learning Engineer
Giga-tech
- Developed data preparation, training, and inference pipelines with FastAPI and PyPI package for shallow parsing, achieving state-of-the-art results with a 96.47 F-score on the Penn Treebank.
- Conducted transformer-based model experimentation, cross-dataset analysis, and statistical significance tests for sentiment analysis, leading to a publication as the first author at KDD 2023.
- Achieved state-of-the-art results in named entity recognition on all Bangla datasets, published in NAACL 2025.
- Developed a Bangla lemmatizer with 98.17% accuracy, published in EMNLP 2023.
Project History
The Brief News
https://the-brief.news/Key Features
• Built a semantic deduplication engine to identify the same news stories across multiple news sources, utilizing Gemini to generate unified and unbiased summaries.
• Developed an intelligent news assistant using LangGraph and vector databases to perform agentic RAG, allowing users to query the corpus for deep analytical insights.
• Engineered automated visualization tools within the assistant to extract data-driven trends and generate real-time charts from the news database.
Analytics Agentic AI Assistant
https://www.iqvia.com/solutions/innovative-models/artificial-intelligence-and-machine-learning/iqvia-ai-assistantThe platform leverages fine-tuned foundation models, advanced prompt engineering, and AI-driven analytics to generate insights from high-volume, complex datasets while maintaining low latency and cost efficiency. Built with a modular, scalable design, it supports secure cloud deployment and seamless integration with existing enterprise data platforms.
This solution enables organizations to automate knowledge-intensive workflows, perform advanced big data analytics conversationally, and accelerate decision-making through reliable, production-ready generative AI.
Bangla NLP Ecosystem Development
https://corpus.bangla.gov.bd/• Led the end-to-end development of shallow and deep parsing systems, including data preparation, separate training pipelines, and a joint modeling framework enabling simultaneous optimization of both tasks.
• Engineered production-ready inference pipelines with a PyPI package, FastAPI service, and Dockerized deployment.
• Achieved state-of-the-art performance on the Penn Treebank benchmark:
– Shallow Parsing: 96.47 F-score
– Deep Parsing: 97.49% UAS and 96.36% LAS
• Developed BanLemma, a Bangla rule-based lemmatizer through collaboration with linguists, achieving 98.17% accuracy on internal evaluation and resulting in an EMNLP submission.
• Built baseline and advanced Bangla NER systems using BERT, BiLSTM, and CRF architectures, incorporating class imbalance strategies and gazetteer integration to improve macro F1 beyond previous SOTA.
• Prepared pre-training boilerplates for BERT, XLNet, and Word2Vec using Hugging Face Transformers and Datasets.
• Developed a Romanized Bangla classifier using CNN-based architectures to support the creation of a corpus.
Human Monitoring & Face Anti-spoofing System
• Developed a live monitoring system for office lobby cameras, enabling person detection, tracking, and recognition in production environments.
• Utilized YOLOv8 for robust person detection and DeepSORT for multi-object tracking across video streams.
• Implemented ArcFace-based facial embedding extraction with cosine similarity matching for accurate face recognition.
• Built face anti-spoofing pipelines using behavioral and liveness cues, including smile detection, eye-blink verification, hand gesture analysis, and head movement validation.
• Optimized the system for real-time inference and practical deployment scenarios requiring high reliability and low latency.
Bangla Next Sequence Prediction
https://ieeexplore.ieee.org/document/9333518Proposed a solution using a Trie and a combination of LSTM and N-gram to predict the relevant next sequence list in Bangla, published in IEE (ICAICT '20 ).
Handwritten Character Image Generation
Education
Bachelor of Science Degree in Computer Science and Engineering
Shahjalal University of Science and Technology - Sylhet, Bangladesh
Certifications
Neural Networks and Deep Learning
Coursera
Skills
Libraries/APIs
PyTorch, TensorFlow, Scikit-learn, Pandas, Matplotlib, Claude API, Hugging Face Transformers, OpenAI API, REST APIs, Pydantic, NumPy, SQLAlchemy, OpenCV
Tools
Claude, Git, ChatGPT, Claude Agent SDK, GitHub, Amazon SageMaker
Languages
Python, SQL, Snowflake
Frameworks
LangGraph, Agentic Frameworks, Streamlit
Platforms
Jupyter Notebook, Docker, LangSmith, Amazon Web Services (AWS), Google Cloud Platform (GCP)
Storage
JSON, Data Pipelines, MongoDB, Amazon S3 (AWS S3), Elasticsearch, PostgreSQL
Paradigms
Microservices, Model Context Protocol (MCP), Object-oriented Programming (OOP)
Other
FastAPI, Machine Learning, Deep Learning, Natural Language Processing (NLP), LangChain, Amazon Bedrock AgentCore, AI Agents, Large Language Models (LLMs), Artificial Intelligence (AI), Retrieval-augmented Generation (RAG), Agentic AI, Generative Artificial Intelligence (GenAI), Prompt Engineering, Deep Neural Networks (DNNs), Hugging Face, Data Science, Anthropic, AI Chatbots, AI Development, Model Deployment, Agentic RAG Systems, Machine Learning Operations (MLOps), Deployment, Open-source LLMs, Chatbots, Local Hosting, Chatbot Conversation Design, AI Assistants, AI Modeling, AI Integration, ChatGPT API, ChatGPT Prompts, AI Pipeline, Large Language Model Operations (LLMOps), Qdrant, Transformers, Data Quality, Cursor AI, ML Pipelines, Multi-agent Systems, Conversational AI, ChromaDB, Generative Adversarial Networks (GANs), Research, APIs, Fine-tuning, Small Language Models (SLMs), Vector Databases, Architecture, Product Development, Cloud Services, Containers, Vector Stores, Data Protection, OpenAI, GPU Computing, MLflow, Image Generation, Human-in-the-loop (HITL), Text to Image, RAG Architecture, RAG Pipelines, Computer Vision, Software Development Lifecycle (SDLC), Digital Signal Processing, Long Short-term Memory (LSTM), N-gram Language Models, Semantic Search, YOLOv5, DeepSORT
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
Share your needs
Choose your talent
Start your risk-free talent trial
Top talent is in high demand.
Start hiring