Muntasir Hossain
I am a data scientist with expertise in big data analysis, machine learning (ML) and deep learning, computational modelling, predictive modelling, computer vision and generative AI. I have a proven track record of delivering impactful results in diverse areas such as energy, technology, and cybersecurity. I have practical experience in developing end-to-end machine learning workflow including data preprocessing, model training at scale and model evaluation, deploying in production and model monitoring with data pipeline automation.
View my LinkedIn profile
Selected projects in data science, machine learning, deep learning, and LLMs.
Neural Network-Based Time-Series Forecasting (CNN-LSTM)
This project implements a multi-step time-series forecasting model using a hybrid CNN-LSTM architecture. The 1D convolutional neural network (CNN) extracts spatial features (e.g., local fluctuations) from the input sequence, while the LSTM network captures long-term temporal dependencies. Unlike recursive single-step prediction, the model performs direct multi-step forecasting (Seq2Seq), outputting am entire future sequence of values at once. Trained on historical energy data, the model forecasts weekly energy consumption over a consecutive 10-week horizon, achieving a Mean Absolute Percentage Error (MAPE) of 10% (equivalent to an overall accuracy of 90%). The results demonstrate robust performance for long-range forecasting, highlighting the effectiveness of combining CNNs for feature extraction and LSTMs for sequential modeling in energy demand prediction.
Figure: Actual and predicted energy usage over 10 weeks of time period.
MLOps with AWS: End-to-End ML Pipelines and Deploymen
Develop an end-to-end machine learning (ML) workflow with automation for all the steps including data preprocessing, training models at scale with distributed computing (GPUs/CPUs), model evaluation, deploying in production, model monitoring and drift detection with Amazon SageMaker Pipeline - a purpose-built CI/CD service.
Figure: ML orchestration reference architecture with AWS
Figure: CI/CD pipeline with Amazon Sagemaker
Fine-tuning LLMs with ORPO & QLoRA (Mistral-v0.3)
ORPO (Odds Ratio Preference Optimization) is a single-stage fine-tuning method to align LLMs with human preferences efficiently while preserving general performance and avoiding multi-stage training. This method trains directly on human preference pairs (chosen, rejected) without a reward model or reinforcement learning (RL) loop, reducing training complexity and resource usage. However, fine-tuning an LLM (e.g. full fine-tuning) for a particular task can still be computationally intensive as it involves updating all the LLM model parameters. Parameter-efficient fine-tuning (PEFT) updates only a small subset of parameters, allowing LLM fine-tuning with limited resources. Here, I have fine-tuned the Mistral-7B-v0.3 foundation model with ORPO and QLoRA (a form of PEFT), by using NVIDIA L4 GPUs. In QLoRA, the pre-trained model weights are first quantized with 4-bit NormalFloat (NF4). The original model weights are frozen while trainable low-rank decomposition weight matrices are introduced and modified during the fine-tuning process, allowing for memory-efficient fine-tuning of the LLM without the need to retrain the entire model from scratch.
Check the model on Hugging Face hub!
Evaluating Safety and Vulnerabilities of LLM apps
Overview
This project demonstrates iterative red-teaming of a policy assistant designed to answer questions about a government-style digital services policy, while strictly avoiding legal advice, speculation, or guidance on bypassing safeguards. The focus is on safety evaluation, failure analysis, and mitigation, rather than model fine-tuning.
Model Separation Strategy
The system intentionally uses different models for generation and evaluation:
- Query responses are generated using gpt-4o-mini
- Safety evaluation is performed using gpt-4o via Giskard detectors This reflects common red-teaming practice: lighter models are sufficient for generation, while stronger models provide more reliable safety judgments. Separating generation and evaluation also avoids self-evaluation effects and keeps evaluation costs controlled.
Initial Evaluation
The policy assistant was evaluated using Giskard across prompt-injection, misuse, and bias detectors. The scan identified multiple failures where the agent did not attempt to answer questions based on the provided policy document. These were not hallucinations or unsafe outputs, but overly conservative refusals.
Figure 1: Initial scan results from Giskard.
Analysis
The root cause was over-refusal. The safety layer correctly blocked requests involving legal advice, speculation, or bypassing safeguards, but also refused some benign questions that could have been partially answered using neutral policy language. This reduced policy grounding and triggered Giskard failures.
Mitigation
The refusal strategy was refined to better distinguish between:
- questions requiring refusal, and
- questions that can be answered safely using policy text alone. Refusals were standardized using fixed, auditable messages, while benign queries now trigger policy-based responses where possible. Safety guarantees were preserved.
Outcome
A follow-up Giskard scan showed improved behavior:
- fewer false positives for “did not attempt to answer”
- stronger grounding in policy text
- no regression in prompt-injection or misuse resistance
Figure 2: Post mitigation scan results from Giskard.
This project demonstrates a complete red-teaming loop — evaluation, failure analysis, mitigation, and re-evaluation — and shows how safety behavior can be systematically improved without increasing risk or cost.
View project and source codes on GitHub
Retrieval-Augmented Generation (RAG) with LLMs and Vector Databases
RAG is a technique that combines a retriever and a generative LLM to deliver accurate responses to queries. It involves retrieving relevant information from a large corpus and then generating contextually appropriate responses to queries. Here, I used the open-source Llama 3 and Mistral v2 models and LangChain with GPU acceleration to perform generative question-answering (QA) with RAG.
View example codes for introduction to RAG on GitHub
Try my app below that uses the Llama 3/Mistral v2 models and FAISS vector store for RAG on your PDF documents!
Analysis & Interactive Visualisation of Global CO₂ Emissions
The World Bank provides data for greenhouse gas emissions in million metric tons of CO₂ equivalent (Mt CO₂e) based on the AR5 global warming potential (GWP). It provides information on environmental impact at both national, regional and economic levels over the past six decades.
Analytical approach:
Time-series aggregation and normalisation across countries, regions, and income groups; comparative cohort analysis across geographic and economic categories; interactive filtering and visual exploration to support exploratory analysis and pattern discovery.
Some key insights from the data:
- Many countries (e.g. China, India) and regions show a clear upward trend in carbon dioxide emissions from 1960 to 2024. For instance, although the population of China increased from 0.82 billion in 1970 to only 1.41 billion in 2023, emission drastically increased from 909 Mt CO₂e to over 13000 Mt CO₂e, reflecting rapid industrialization. While the population of China and India in 2023 were nearly the same, the CO₂ emission from China was 4.5 fold to that of India.
- There is significant variation between countries. Highly industrialized or resource-rich nations (e.g., Saudi Arabia, United Arab Emirates) emit far more CO₂ than smaller or less industrialized countries (e.g., Aruba, Burundi).
- The data suggests a strong link between economic development and emissions growth. Countries experiencing rapid economic expansion (e.g., Vietnam, United Arab Emirates) show marked increases in emissions, while some developed countries (e.g., Germany, Austria, Belgium) have stabilized or slightly reduced their emissions in recent years, likely due to policy interventions or shifts to cleaner energy.
- While ‘High Income’ regions dominated the CO2 emissions pre-2020, the ‘Middle Income’ and ‘Upper Middle Income’ regions rapidly increased CO2 emissions after 2000, exceeding the emissions from ‘High Income’ regions.
Global CO₂ emissions
Figure: Interactive visualization of global CO₂ emissions by country and year
Time sereis CO₂ emissions
Figure: Time sereis CO₂ emissions for selected countries
Population Growth
Figure: Population growth for selected countries
CO₂ emissions by income groups
Figure: Interactive visualization of CO₂ emissions for different income zones from 1970 to 2023
CO₂ emissions by geographic regions
Figure: Interactive visualization of CO₂ emissions for different geographic regions from 1970 to 2023
Computer Vision: Building and Deploying YOLOv8 models for object detection at scale
Deployed a state-of-the-art YOLOv8 object detection model to real-time Amazon SageMaker endpoints, enabling scalable, low-latency inference for image and video inputs. Focused on model serving, endpoint configuration, and operational inference rather than model training.
Figure: Object detection with YOLOv8 model deployed to a real-time Amazon SageMaker endpoints.
