I build data-centric, explainable, and interactive AI systems, from problem framing and data pipelines to model evaluation, dashboards, APIs, and decision-support tools.
My work focuses on projects where machine learning is not only trained, but also tested, explained, documented, and connected to a real user or business decision.
Main focus areas:
- Retrieval-Augmented Generation and graph-based AI systems
- Risk modeling, fraud detection, underwriting, and decision safety
- NLP systems with responsible evaluation and uncertainty handling
- Synthetic data generation and realism evaluation
- SQL + Python machine learning workflows
- Streamlit/FastAPI apps for interactive model use
Open to: Data Scientist, Machine Learning Engineer, and Applied AI roles, especially projects involving explainability, decision support, and production-minded ML workflows.
| Project | Area | Why it matters |
|---|---|---|
| Fake-News-Detector | NLP / Responsible AI | TF-IDF + Logistic Regression style-risk detector with Streamlit, CLI prediction, uncertainty handling, leakage analysis, tests, and CI. |
| Coffee-Shop-Profit-Predictor | SQL + Machine Learning | End-to-end site-selection workflow with SQL feature engineering, regression modeling, model comparison, candidate ranking, tests, and CI. |
| Graph-RAG-Engine | RAG / LLM Systems | Explainable Graph + Vector + RAG system with FAISS retrieval, knowledge graph reasoning paths, FastAPI backend, and Streamlit UI. |
| Synthetic-Data-Artist | Synthetic Data / Generative ML | Research-style comparison of Gaussian Copula and VAE methods with distribution checks, correlation analysis, PCA diagnostics, and visual reports. |
| Financial-Fraud-Risk-Engine | Fraud Detection / Risk ML | Cost-sensitive fraud detection system with SHAP explanations, threshold optimization, batch scoring, and an interactive dashboard. |
| Underwriting-Decision-Safety-Lab | Decision Safety / Risk ML | Loan approval safety lab with probability calibration, abstention policies, coverage-quality tradeoffs, triage UI, and data quality checks. |
Projects that move beyond notebooks into repeatable workflows:
- data cleaning and validation
- feature engineering
- train/test evaluation
- cross-validation and model comparison
- uncertainty and threshold handling
- saved artifacts and reproducible outputs
- tests and CI where appropriate
I like building model interfaces that a user can actually interact with:
- Streamlit dashboards
- FastAPI backends
- CLI tools
- batch scoring workflows
- visual reports
- decision-support outputs
I try to make model behavior understandable through:
- honest limitations
- model cards and documentation
- leakage analysis
- SHAP and feature importance
- calibration and abstention
- uncertainty bands
- human-review workflows
- Graph-RAG-Engine | Graph intelligence, vector search, and explainable retrieval-augmented generation.
- PR-Guardian-AI | AI-assisted pull request review workflow using GitHub App patterns and OpenAI integration.
- RAG-vs-Fine-Tuning | Practical framework for deciding between retrieval and fine-tuning.
- Designing-Hybrid-AI-Systems | Notes and patterns for hybrid AI design.
- Financial-Fraud-Risk-Engine | Fraud-risk workflow with cost-sensitive thresholding and SHAP explanations.
- Underwriting-Decision-Safety-Lab | Calibration, abstention, and defensible loan decision policies.
- Onchain-Security-Suite | Web3 security pipeline with static analysis and deployer risk scoring.
- Synthetic-Data-Artist | Copula vs VAE synthetic tabular data comparison and diagnostics.
- Autocurator-Synthetic-Data-Benchmark | Synthetic data benchmarking across fidelity, coverage, privacy, and utility.
- Missing-Data-Doctor | Missingness profiling and imputation impact analysis.
- Coffee-Shop-Profit-Predictor | SQL + ML workflow for retail location profitability prediction.
- Forecast-Factory | Forecasting and scenario simulation app for business decision support.
- Data-Storytelling-Dashboard | E-commerce analytics dashboard with KPIs, cohorts, retention, and business narrative.
- Market-IQ | BI-style analytics and KPI exploration.
- Fake-News-Detector | Responsible text classification with uncertainty handling and leakage analysis.
- Sentiment-Analysis-BERT | BERT fine-tuning pipeline for sentiment classification.
- ML-Playground-Autodetect | Interactive ML playground with automatic task detection.
| Area | Tools |
|---|---|
| Languages | Python, SQL, Solidity, MQL5 |
| Data | pandas, NumPy, SQLite, SQLAlchemy |
| Machine Learning | scikit-learn, XGBoost, LightGBM, joblib |
| Deep Learning / NLP | PyTorch, TensorFlow/Keras, Hugging Face Transformers, BERT |
| RAG / AI Systems | FAISS, Sentence Transformers, FastAPI, Streamlit |
| Visualization | matplotlib, Plotly, Streamlit dashboards |
| Explainability | SHAP, feature importance, calibration, threshold analysis |
| Workflow Quality | tests, CI, reproducible outputs, model artifacts, documentation |
- Building more production-style ML projects with tests, CI, and reproducible workflows.
- Improving RAG systems with better retrieval evaluation, traceability, and source grounding.
- Expanding risk and decision-safety projects with calibration, abstention, and monitoring ideas.
- Turning analytics projects into clearer decision-support tools with stronger documentation and outputs.
- Email: amirhosseinhonardoust@gmail.com
- LinkedIn: honardoust
- X: @amirhonardoust
AI is not just about models; it is about systems that solve real problems for real people.