Data Scientist & Machine Learning Engineer
I design data-centric, explainable, and interactive AI systems.

All Repositories • LinkedIn • Email

• Data Scientist & ML Engineer with a Physics background and hands-on experience across:
  • Time series & forecasting
  • NLP and text classification
  • Computer vision & recommendation systems
  • Algorithmic trading & risk modeling
• Comfortable with the full ML lifecycle:
  • Problem framing → data generation/collection → feature engineering → modeling → evaluation → deployment (Streamlit/FastAPI dashboards).
• Strong focus on:
  • Synthetic data & robustness
  • Interactive dashboards & decision-support tools
  • LLMs, RAG & hybrid AI architectures
  • Explainability & human-centered AI
• I enjoy writing deep, structured explanations, many repos are part code, part essay.

Languages & Tools

• Languages: Python, SQL, MQL5, a bit of Solidity
• ML / DL: PyTorch, TensorFlow/Keras, scikit-learn, XGBoost, LightGBM
• Time Series & Forecasting: ARIMA/SARIMA, Prophet, LSTMs, rolling windows
• NLP: BERT, TF-IDF, classic ML (LogReg, Naive Bayes, SVM)
• Apps & Services: Streamlit, FastAPI, Plotly, SQLite/SQLAlchemy
• MLOps-ish / Analysis: SHAP, feature importance, evaluation frameworks, synthetic data benchmarks

Domains & Topics

• Synthetic & tabular data quality
• Forecasting & scenario simulation
• Recommender systems
• Smart contract risk analytics
• Interactive data storytelling
• LLMs & Retrieval-Augmented Generation (RAG)
• Algorithmic trading & financial modeling

I organize my projects into a few clusters so you can jump directly to what interests you:

1. Synthetic Data, Data Realism & Anomaly Detection
2. Forecasting, Dashboards & Interactive ML
3. LLMs, RAG & Hybrid AI Systems
4. Core ML, Deep Learning & Portfolio Projects
5. Explainability, Thinking Like a Data Scientist & AI Philosophy
6. Smart Contracts, Security & Risk Analytics

Generating realistic tables, probing data “authenticity”, and stress-testing models.

These projects focus on fidelity, coverage, privacy, and utility of synthetic data, plus anomaly detection in tabular domains.

• Autocurator-Synthetic-Data-Benchmark
  A benchmarking toolkit for synthetic tabular data generators:

  • Compares different models (GANs, VAEs, copulas, etc.)
  • Evaluates distribution fidelity, feature coverage, privacy leakage, and downstream ML utility
  • Produces visual reports (PCA, correlations, histograms) to understand where generators succeed or fail
    Goal: make it easier to choose the right synthetic data approach for a business use case.

• Synthetic-Data-Artist
  Deep dive into Gaussian Copula vs VAE for tabular data:

  • Side-by-side comparison of marginal and joint distributions
  • PCA visualizations of real vs synthetic embeddings
  • Correlation matrix similarity, pair plots, and coverage analysis
    Think of it as a “microscope” for synthetic tabular data.

• Anomaly-Detection
  End-to-end anomaly detection on synthetic transactions/sales:

  • Data generation with realistic “weird” patterns injected
  • Uses Isolation Forest, Local Outlier Factor, and classic statistical methods
  • Visual diagnostics and confusion-matrix-style evaluations
    Useful for building intuition about anomalies in finance/ops data.

• Market-Basket-Analysis
  Retail-style synthetic purchase data:

  • Apriori & FP-Growth frequent itemset mining
  • Association rules with support, confidence, and lift
  • Exportable rules + quick visual summaries
    Foundation for recommendation, cross-sell, and promo design.

• Sales-Data-Analysis
  Lightweight but complete:

  • Synthetic sales dataset generation
  • Cleaning, aggregation, KPI dashboards
  • Time-based trend analysis and segmentation
    Great for explaining analytics pipelines to non-technical stakeholders.

• Missing-Data-Doctor
  Toolkit for missingness profiling & imputation:

  • Visual missingness maps and patterns (by column, row, time)
  • Simple and advanced imputation strategies
  • Before/after comparisons for ML performance
    Focus: understanding how missing data distorts models.

• Noise-Injection-Techniques
  Experiments on robustness via controlled noise:

  • Add noise to tabular features/labels during training
  • Explore how different noise types affect generalization
  • PyTorch-based training loops and results visualization
    Bridge between data augmentation and robustness in non-vision domains.

Treat forecasting and analytics as interactive tools, not static reports.

These projects focus on Streamlit dashboards, scenario analysis, and business-friendly UIs.

• Forecast-Factory
  Forecasting & simulation app:

  • Streamlit UI to upload time series (sales, traffic, revenue)
  • Uses Prophet (and/or other models) for forecasting with confidence intervals
  • Lets users run “what if we change X?” simulations on key drivers
    Designed for business teams to explore future scenarios without touching code.

• Market-IQ
  BI-style web app:

  • Ingests transactional/sales-like data
  • Computes core KPIs (revenue, retention, AOV, etc.)
  • Time-series charts, comparisons, and exportable reports
    Acts like a focused analytics tool for small/medium businesses.

• Data-Storytelling-Dashboard
  End-to-end narrative dashboard:

  • E-commerce style dataset with customers, orders, and products
  • KPIs, cohort analysis, and retention curves
  • Visuals + narrative “takeaways” to interpret the charts
    Focuses on storytelling, not just plotting.

• Beyond-Charts-Interactive-Storytelling
  Code + essay:

  • RFM segmentation, cohort tracking, user lifecycle
  • Interactive views that adapt to user selections
  • Conceptual guide on how to build narrative dashboards
    For people who want to turn dashboards into decision tools.

• AI-Report-Factory
  Automated reporting:

  • Input: structured data + configuration
  • Output: KPIs, visualizations, and narrative sections in Markdown/HTML
  • Uses templating to make the reporting repeatable
    Ideal for recurring reports that still need a “human-readable” style.

• AI-Personal-Study-Tracker
  Productivity & study analytics:

  • Streamlit interface for logging study sessions
  • SQLite backend for persistence
  • ML model (RandomForestRegressor) to predict productivity and surface patterns
    Example of sending ML back to the user as personal feedback.

• Demand-Forecasting
  Classic time-series pipeline:

  • Synthetic demand & seasonality
  • ARIMA/SARIMA modeling workflow
  • Forecast evaluation and plots
    Template for demand planning and inventory decisions.

• ML-Playground-Autodetect
  Auto ML playground:

  • Streamlit UI where you upload a dataset
  • Automatically detects classification vs regression
  • Builds sensible ML pipelines + evaluation
    Useful for teaching and quick sanity checks.

Building explainable, data-grounded LLM systems with retrieval & graphs.

• Graph-RAG-Engine
  A more structured take on RAG:

  • Vector search (FAISS) for semantic retrieval
  • Knowledge graph to add structure and relationships
  • FastAPI backend and optional Streamlit front-end
  • Emphasis on traceability and explaining why an answer was given
    Great for recommendation, research assistants, or domain-specific QA.

• Designing-Hybrid-AI-Systems
  Conceptual + practical:

  • How to combine vector search, knowledge graphs, and LLMs
  • Design patterns for hybrid intelligence
  • Notes on failure modes and interpretability
    A “systems thinking” view for building LLM-powered apps.

• RAG-vs-Fine-Tuning
  Decision framework:

  • When to use RAG, when to fine-tune, when to do both
  • Cost, latency, maintenance, and data constraints
  • Includes examples and architectural diagrams (where applicable)
    Helpful for teams deciding how to productionize LLMs.

Classic ML projects done with clean structure and clear evaluation.

• Stock-LSTM-Forecasting
  Time-series forecasting with LSTMs:

  • Data preparation with sliding windows
  • PyTorch LSTM architecture
  • Loss curves + forecast vs actual plots
    Use case: financial time series, sensor data, or demand.

• Image-Captioning-CNN-LSTM
  Computer vision + language:

  • Pretrained ResNet as image encoder
  • LSTM decoder generating captions word by word
  • BLEU score and qualitative examples
    Classic example of multimodal ML.

• Sentiment-Analysis-BERT
  Transformer-based text classification:

  • Fine-tuning BERT on sentiment data (tweets)
  • Training/evaluation pipeline
  • Confusion matrix, ROC curves, and example predictions
    Template for other classification tasks with BERT-style models.

• Sentiment-Analysis-NLP
  Classical ML for text:

  • Tokenization, stopword removal, lemmatization
  • TF-IDF vectorization
  • Models: Logistic Regression, Naive Bayes, Random Forest
    Shows how far you can go with “non-deep” NLP.

• Movie-Recommendation-System
  Hybrid recommender:

  • Content-based filtering (TF-IDF + cosine similarity)
  • Collaborative filtering via matrix factorization
  • Evaluation with ranking metrics and examples
    Useful base for product/content recommendations.

• LSTM-Time-Series-Forecasting
  Generic LSTM-forecast template:

  • Works on many univariate series
  • Clear code structure and visualizations
    Good starting point for experimenting with sequence models.

• Handwritten-Digit-GAN
  Generative modeling:

  • DCGAN on MNIST
  • Training loop, generated samples, and latent space interpolation
    Intro to generative models in vision.

Code + essays about how we think about and interpret AI systems.

• Shap-Mini
  Minimal SHAP explainability demo:

  • Tabular ML model (tree-based)
  • Global and local SHAP plots
  • Good for teaching “why did the model decide this?”
    Brings explainability down to a small, digestible example.

• Think-Like-a-Data-Scientist
  Long-form essay:

  • Framing questions, hypotheses, and experiments
  • How to move from raw data → insight → action
  • Balancing rigor with storytelling
    More about mindset than code.

• Forecasting-The-Future-of-Forecasting
  Strategic perspective:

  • How forecasting tools shape decisions
  • Interactive foresight vs static point estimates
  • Reflections on feedback loops and reflexive systems

• The-Future-of-Interactive-ML
  Why interactivity matters:

  • Benefits of human-in-the-loop ML
  • Examples with Streamlit-style interfaces
  • How UI/UX changes modeling choices

• Algorithmic-Empath-Human-Fallibility
  Ethics & “algorithmic empathy”:

  • Modeling human mistakes, uncertainty, and disagreement
  • Thinking beyond accuracy: fairness, robustness, trust
    Explores what it means for algorithms to “understand” humans.

• Measuring-The-Soul-of-Data
  Philosophy of data realism:

  • What makes data feel “alive” or “authentic”
  • Exploring relationships, diversity, and subtle patterns
  • Especially in the context of synthetic vs organic data

• Quiet-Machines-Minimalist-AI
  Minimalist AI:

  • Preference for quiet, non-intrusive, human-respecting AI systems
  • Thoughts on attention, overload, and calm technology

Applying ML & static analysis ideas to smart contracts and risk.

• Python-Solidity-Feature-Engineering
  Feature extraction for Solidity contracts:

  • Parse Solidity code using Python tooling
  • Extract structural & semantic features (complexity, patterns)
  • Basis for risk models, security classification, or audit support

• Smart-Contract-Risk-Analyzer
  Static analysis plus heuristics:

  • Identify risky patterns in smart contracts
  • Compute risk scores or categories
  • Designed as a step toward automated security triage

If you’re just browsing and want a quick sense of my work, start here:

• Graph-RAG-Engine
• Forecast-Factory
• Synthetic-Data-Artist
• Data-Storytelling-Dashboard
• Sentiment-Analysis-BERT
• Think-Like-a-Data-Scientist

“AI is not just about models; it’s about systems that solve real problems for real people.”