Fadhil Rusydi HafizhData Analyst
Menu
HomeWorkEvidenceCVContact
Work

Projects built around real analytical problems.

Each case study separates what is live, what was tested, what the result means, and what is still in progress.

All case studies

Project explorer

Filter by recruiter-relevant focus. One project can match more than one focus while its primary category and evidence state stay visible.

Showing all 5 case studies.

Risk analyticsHistorical evidence live

Ruang Risiko IDX

Indonesian equity risk research that makes returns, volatility, drawdown, downside estimates, direction probabilities, provenance, and model evidence inspectable.

FocusData quality and reliability · Modeling and validation · Forecasting · Market and risk research · Product delivery

StackNext.js · Vercel · yfinance · GARCH family models

Built
A production research interface for five IDX equities plus IHSG, backed by validated market artifacts and archived out-of-sample model evidence.
Strongest result
9,072 successful archived walk-forward forecast rows with 0 failed archived rows.
Quick project context
Problem
A price forecast alone does not describe risk. Recruiters and users need to see uncertainty, data quality, benchmark context, and how a model behaved outside its fitting window.
My role
I designed and built the reviewed research workflow across market-data validation, quarantine rules, volatility and direction modeling, out-of-sample evaluation, and the public research interface.
Result
9,072 successful archived walk-forward forecast rows with 0 failed archived rows.
Stack
Next.js · Vercel · yfinance · GARCH family models · scikit-learn
Status
Historical evidence live
Read case studyOpen live product
Risk analyticsVerified 2026-09-07
Ruang Risiko IDX
36archived model combinations
9,072walk-forward rows
0failed archived rows
Project snapshot graphic, not a product screenshot
Market intelligenceResearch prototype

Cakrawala Intelligence Terminal

Forex-first market and macro research that makes source health, evidence freshness, model state, and decision boundaries visible.

FocusData quality and reliability · Market and risk research · Product delivery

StackPython · Streamlit · market data adapters · model governance

Built
A public research terminal across 12 analytical areas, with guarded data ingestion and explicit separation between public research and owner-only analysis.
Strongest result
Recorded product scope covers 12 analytical areas with daily research and weekly model-governance design.
Quick project context
Problem
A market conclusion is weak when the evidence trail is hidden. Provider reliability, point-in-time evidence, authorization, and model state have to be treated as product requirements.
My role
I designed the research terminal, provider and evidence boundaries, model-governance rules, and the separation between public research and owner-only workflows in the reviewed scope.
Result
Recorded product scope covers 12 analytical areas with daily research and weekly model-governance design.
Stack
Python · Streamlit · market data adapters · model governance
Status
Research prototype
Read case study
Market intelligenceVerified 2026-08-29
Cakrawala Intelligence Terminal
12public analytical areas
Dailyresearch cadence
Weeklymodel governance
Project snapshot graphic, not a product screenshot
Applied machine learningRebuild tested

SpamShield ML

Cost-aware SMS spam classification where probability quality, threshold policy, artifact integrity, and deployment are evaluated as separate decisions.

FocusData quality and reliability · Modeling and validation · Product delivery

StackPython · scikit-learn · FastAPI · Docker

Built
A reproducible UCI training path with TF-IDF, model comparison, threshold analysis, integrity-checked artifacts, FastAPI, Docker, and Streamlit delivery boundaries.
Strongest result
Validation F1 0.9291 with average precision 0.9788 on the fixed stratified rebuild holdout.
Quick project context
Problem
A good classifier can still make the wrong operating decision. A false positive can hide a legitimate message, so threshold policy should reflect an explicit cost assumption rather than defaulting silently to 0.50.
My role
I rebuilt and evaluated the UCI classification workflow, compared models and thresholds, defined the artifact contract, and designed the FastAPI, Docker, and Streamlit delivery boundaries.
Result
Validation F1 0.9291 with average precision 0.9788 on the fixed stratified rebuild holdout.
Stack
Python · scikit-learn · FastAPI · Docker · Streamlit · TF-IDF
Status
Rebuild tested
Read case study
Applied machine learningVerified 2026-08-29
SpamShield ML
0.9291validation F1
0.9788average precision
0.0347Brier score
Project snapshot graphic, not a product screenshot
Data engineeringLive dashboard

PanganLens Indonesia

Public-benefit food-price analytics designed around provenance, idempotent loading, conflict handling, revision history, and warehouse quality.

FocusData quality and reliability

StackPIHPS Bank Indonesia · BigQuery · 3NF data modeling · data quality controls

Built
A public dashboard deployment plus an evidence-first data architecture direction for PIHPS food-price acquisition, normalized BigQuery storage, revision handling, quarantine, and curated reporting.
Strongest result
The public production route is live, while live acquisition, warehouse coverage, and current data behavior remain separate verification gates.
Quick project context
Problem
A food-price dashboard is only trustworthy when every observation has an inspectable source, business identity, quality decision, and revision path.
My role
I am designing the acquisition, warehouse, data-quality, revision, and reporting architecture. Live acquisition remains the next verification gate, so no production coverage claim is made yet.
Result
The public production route is live, while live acquisition, warehouse coverage, and current data behavior remain separate verification gates.
Stack
PIHPS Bank Indonesia · BigQuery · 3NF data modeling · data quality controls
Status
Live dashboard
Read case studyOpen live product
Data engineeringVerified 2026-09-07
PanganLens Indonesia
3NFwarehouse core
18:00 WIBplanned refresh
Fail closedquality stance
Project snapshot graphic, not a product screenshot
ForecastingArchived project

DSC MCF ITB 2025

Claims forecasting where leakage control, lag correctness, and chronological evaluation matter more than adding another model.

FocusModeling and validation · Forecasting

StackLightGBM · Prophet · Holt-Winters ETS · time-series validation

Built
A historical LightGBM, Prophet, and Holt-Winters forecasting workflow, followed by a review that identified three material temporal-pipeline defects.
Strongest result
Historical MAPE 4.08%, retained as a historical result until a leakage-safe rebuild proves comparable validity.
Quick project context
Problem
Time-series accuracy is meaningless if a pipeline can see the future or if lag logic silently changes which periods are evaluated.
My role
I reviewed the historical forecasting workflow and identified target-equivalent leakage, lag-propagation defects, and row loss that must be removed before a rebuild result can be compared fairly.
Result
Historical MAPE 4.08%, retained as a historical result until a leakage-safe rebuild proves comparable validity.
Stack
LightGBM · Prophet · Holt-Winters ETS · time-series validation
Status
Archived project
Read case study
ForecastingVerified 2026-08-29
DSC MCF ITB 2025
4.08%historical MAPE
3material pipeline defects
3model families
Project snapshot graphic, not a product screenshot
Verified project screenshot

Open verified source