Hi, I'm Md. Alif Hossen

Entry-Level Machine Learning Engineer · Applied AI Developer · Data Scientist

I build practical AI systems that move from data to decisions.

CSE graduate building end-to-end solutions across RAG, fraud risk, demand forecasting, computer vision, and analytics—with evaluation, APIs, testing, monitoring, and deployment in the loop.

📍 Dhaka, Bangladesh · Open to entry-level AI/ML, Data Science, and Data Analytics opportunities

I am a B.Sc. in Computer Science & Engineering graduate from Daffodil International University. My work focuses on building practical machine-learning and applied-AI systems beyond isolated notebooks—through reproducible pipelines, careful validation, APIs, interactive applications, automated tests, and deployable workflows.

Across my projects, I work with retrieval systems, cost-sensitive classification, forecasting, computer vision, explainability, and monitoring concepts. I am seeking an entry-level role where I can contribute to useful AI and data products while continuing to strengthen production ML engineering skills.

Live demos are hosted on Streamlit Community Cloud and may take a few seconds to wake up.

Problem: Financial filings are long and difficult to search reliably while preserving source context.

What I built:

• Citation-grounded research over SEC 10-K and 10-Q filings
• Structure-aware processing with hybrid BM25 + FAISS retrieval, reciprocal-rank fusion, and cross-encoder reranking
• SEC XBRL Company Facts integration and deterministic financial calculations behind FastAPI and Streamlit

Technical highlights: Python · BM25 · FAISS · Sentence Transformers · FastAPI · Streamlit · SQLite · Docker · pytest

Why it matters: It turns complex regulatory filings into a traceable research workflow instead of relying on unsupported model-generated answers.

Problem: Fraud classifiers need more than accuracy—they require time-aware validation, calibrated risk scores, explainability, and decision thresholds.

What I built:

• Fraud-risk pipeline with temporal separation and validation-based model selection
• Probability calibration, cost-sensitive thresholds, and SHAP-based reason codes
• FastAPI inference, Streamlit application, tests, monitoring modules, and optional PostgreSQL persistence

Technical highlights: Python · Logistic Regression · LightGBM · XGBoost · SHAP · FastAPI · Streamlit · PostgreSQL · MLflow · Evidently · Docker · pytest

Why it matters: It connects model predictions to explainable, cost-aware decisions and operational monitoring concepts.

Problem: Demand forecasts are useful only when validated correctly and translated into inventory decisions.

What I built:

• M5-based workflow with leakage-safe lag and rolling features
• Seasonal Naive and LightGBM models with chronological validation and rolling-origin backtesting
• Prediction intervals and inventory recommendations through a FastAPI-backed Streamlit application

Technical highlights: Python · Pandas · LightGBM · Time-Series Validation · FastAPI · Streamlit · Docker · GitHub Actions

Why it matters: It connects forecast uncertainty with practical safety-stock, reorder, and stock-risk decisions.

Problem: Reviewing long road-video recordings manually is slow and makes potential traffic events difficult to organise.

What I built:

• Academic video-screening prototype using YOLO detection, DeepSORT-style tracking, plate-region detection, and EasyOCR
• Annotated video, vehicle crops, summary counts, and CSV exports through Streamlit
• Transparent rule-based outputs designed for review rather than automatic enforcement

Technical highlights: Python · YOLO · OpenCV · DeepSORT · EasyOCR · NumPy · Pandas · Streamlit · pytest

Why it matters: It organises potential events for faster review while keeping human judgement in the loop.

Speed and lane results are heuristic screening outputs that require human review.

flowchart TD
    A["Data"] --> B["Validation & Cleaning"]
    B --> C["Features or Retrieval"]
    C --> D["Model or AI System"]
    D --> E["Evaluation"]
    E --> F["API or Application"]
    F --> G["Deployment & Monitoring"]

My projects focus on moving beyond notebook-only experimentation toward reproducible, testable, and usable AI applications. They are portfolio and academic systems—not claims of enterprise production experience.

• Reproducible workflows: structured projects, configuration, dependencies, and repeatable execution
• Reliable evaluation: temporal splits, baselines, calibration, backtesting, and leakage-aware validation
• Explainable outputs: SHAP reason codes, citations, deterministic calculations, and reviewable results
• Usable delivery: FastAPI inference, Streamlit interfaces, persistence, and containerisation
• Quality controls: automated tests, GitHub workflows, monitoring concepts, and explicit limitations

• Production ML engineering and maintainable service design
• MLOps workflows, experiment tracking, and model monitoring
• Advanced RAG evaluation and retrieval quality
• Scalable ML APIs and reliable deployment practices

These are active learning areas, not claims of professional production expertise.

• Applied AI: grounded retrieval and practical RAG applications
• ML Engineering: validation, explainability, APIs, and testing
• Decision Systems: forecasting, inventory optimisation, and fraud-risk modelling
• Computer Vision: reviewable video-analysis and OCR workflows

GitHub activity and contribution history are available directly on my profile; this avoids unreliable third-party statistics images.

I am open to entry-level opportunities in AI/ML Engineering, Machine Learning, Applied AI, Data Science, and ML-focused Data Analytics.