amir@honardoust.codes:~$ ./open-lab-index

Technical work, organized as experiment records.

A living archive for project reasoning, model evaluation, system design, risk workflows, reproducible experiments, and technical notes behind my data science portfolio.

$ list experiments --status published --tag evaluation

Open lab index Read EXP-001 GitHub

records12

experiments07

notes05

modepublic

{
  "owner": "Amir Honardoust",
  "role": "Data Scientist",
  "format": "experiment archive",
  "main_site": "honardoust.me"
}

01 / LAB INDEX

Browse records like a technical database.

This is the central index. Each row has an ID, status, type, focus area, and a direct source link.

ID Record Type Status Skills proven Open

EXP-001 Underwriting Decision Safety Lab Decision Safety Published Calibration · Abstention · Validation · Slices detail →

EXP-002 Financial Fraud Risk Engine Risk Modeling Published Fraud ML · Thresholds · SHAP · Reason codes detail →

EXP-003 Graph-RAG Engine AI System Published NLP · RAG · Graph retrieval · APIs detail →

EXP-004 Synthetic Data Artist Evaluation Lab Published Copula · VAE · Data quality · Privacy proxy detail →

EXP-005 Movie Recommendation System Recommender System Published TF-IDF · SVD · Baselines · Alpha sweep detail →

EXP-006 Fake News Detector NLP Pipeline Published TF-IDF · Classification · Evaluation · App detail →

EXP-007 Coffee Shop Profit Predictor Business Analytics Published SQL · Regression · CV · Candidate scoring detail →

NOTE-001 Machine Learning Warning Systems Essay Published Human-in-loop · Risk design · ML governance read →

NOTE-002 Coverage Is the Silent Killer Essay Published Data quality · Coverage · KPI reliability read →

NOTE-003 KPI Denominator Truth Essay Published KPI design · Product analytics · Metrics read →

NOTE-004 Abstention Is a Product Feature Essay Published Calibration · Thresholding · Human review read →

NOTE-005 Designing Hybrid AI Systems Essay Published RAG · Knowledge graphs · Explainable AI read →

Flip through selected experiment records.

02 / FEATURED RECORD

EXP-001 · Underwriting Decision Safety Lab

open record →

Question

Can an underwriting model know when to defer uncertain decisions to human review?

Method

Train a probability model, evaluate calibration, select abstention thresholds, validate inputs, and expose review decisions through generated artifacts and a dashboard.

Evidence

Metrics, reliability diagrams, coverage curves, policy variants, slice reports, tests, CI, and generated prediction files.

What it proves

Risk-aware ML design, calibration, selective prediction, governance-minded evaluation, and dashboard communication.

loan data validation probability model calibration abstention slice report review UI

02 / FEATURED RECORD

EXP-002 · Financial Fraud Risk Engine

open record →

Question

How can fraud-risk scores become analyst-review decisions instead of just model outputs?

Method

Generate realistic synthetic transactions, train a fraud-risk pipeline, search cost-sensitive thresholds, produce policy artifacts, and score transactions for analyst review.

Evidence

ROC/PR curves, Brier score, threshold policy files, scored CSVs, reason codes, dashboard helpers, unit tests, and CI smoke workflows.

What it proves

Risk modeling, class-imbalance evaluation, threshold policy thinking, explainability, and operational workflow design.

transactions validation fraud model threshold search policy artifacts reason codes dashboard

02 / FEATURED RECORD

EXP-003 · Graph-RAG Engine

open record →

Question

Can retrieval paths make RAG answers more explainable, grounded, and easier to inspect?

Method

Chunk documents, embed text, build an in-memory knowledge graph, retrieve with vector and graph signals, and expose inspectable answers through API/UI layers.

Evidence

Retrieval metrics, golden-query evaluation, citations, graph paths, API contracts, tests, CI, and optional LLM fallback behavior.

What it proves

Applied AI architecture, retrieval evaluation, system thinking, and explainable AI output design.

documents chunks embeddings knowledge graph hybrid retriever citations answer UI

02 / FEATURED RECORD

EXP-004 · Synthetic Data Artist

open record →

Question

Which synthetic data method better preserves the behavior of real tabular data?

Method

Generate synthetic data with Copula and VAE methods, then compare outputs against real data using statistical, visual, privacy, and ML-utility diagnostics.

Evidence

Distribution overlap, categorical similarity, correlation difference, boundary violations, nearest-neighbor privacy proxy, PCA plots, pairplots, and quality summaries.

What it proves

Model comparison, evaluation design, data quality reasoning, CLI workflows, and transparent reporting.

real data schema detection generator synthetic data quality metrics plots report

02 / FEATURED RECORD

EXP-005 · Movie Recommendation System

open record →

Question

When does a hybrid recommender actually beat simple recommendation baselines?

Method

Generate deterministic synthetic movies and ratings, train content/collaborative/hybrid recommenders, compare against baselines, and tune hybrid alpha by NDCG.

Evidence

Precision@K, Recall@K, NDCG@K, baseline comparison CSV, alpha sweep plot, structured recommendation CSVs, tests, and CI.

What it proves

Recommender evaluation, baseline discipline, reproducibility, metric interpretation, and user-facing recommendation output design.

movies ratings split content model SVD model hybrid sweep recommendations

02 / FEATURED RECORD

EXP-006 · Fake News Detector

open record →

Question

How far can a clean classical NLP pipeline go for fake-vs-real news classification?

Method

Clean article text, vectorize with TF-IDF, train a logistic-regression classifier, evaluate metrics, and expose predictions through a small app.

Evidence

Saved model/vectorizer artifacts, metrics, charts, probability outputs, example predictions, and app behavior.

What it proves

NLP preprocessing, classification, baseline modeling, metric communication, and applied ML app delivery.

news text cleaning TF-IDF classifier metrics artifacts app

02 / FEATURED RECORD

EXP-007 · Coffee Shop Profit Predictor

open record →

Question

Can SQL features and interpretable regression support site-selection decisions?

Method

Build location-level features in SQLite, train regression models, compare against baselines, validate performance, and score candidate locations with risk notes.

Evidence

R²/MAE metrics, cross-validation, baseline comparison, residual plots, feature importance, candidate predictions, tests, and CI.

What it proves

SQL analytics, feature engineering, regression modeling, business communication, and model-risk awareness.

location data SQLite features model comparison validation diagnostics candidate scoring

03 / TECHNICAL ESSAYS

Technical essays from my data science notebook.

These essays capture how I think about reliable data, model behavior, evaluation, metrics, and AI systems beyond the code.

NOTE-001

How I think about data science systems.

01Problem

Define the decision, uncertainty, and success criteria.

02Data

Inspect, clean, transform, validate, and document assumptions.

03Model

Train baselines, compare methods, and evaluate honestly.

04Interface

Expose results through dashboards, APIs, reports, or tools.

05Decision

Communicate limits, tradeoffs, and recommended next actions.

amir@lab:~$ connect

This is my technical lab for experiments, systems, and notes. My main portfolio lives on honardoust.me.

Main portfolio GitHub LinkedIn