DAFFA ZAHY KHAIRY
DAFFA ZAHY KHAIRY

SHOW THENUMBERS.

Explore work Get in touch
Based in Bandung, IndonesiaOpen to business analyst, data and applied AI roles and internships

Selected work

Synthetic or simulated data, tested locally, not deployed. Each case study says what is not proven yet.

Same KPI, three different numbers.

1,848 KPI cells reconciled · 0 mismatchesSynthetic data · tested locally · not deployed

Read the case study

Which customer gets which offer, and does it pay?

NDCG@5 0.818 → 0.911 · 95% CI [+0.079, +0.105]Offline, synthetic · not causal · not deployed

Read the case study

The model has no price field.

0 of 60 briefs sent personal data to the modelSynthetic data · 101 backend tests · runs locally

Read the case study

Approval is where most tools stop.

AI extraction: 1 / 36, then 21 / 50 held out · gate failsSynthetic data · rule-based baseline · runs locally

Read the case study

More built work.

Who will leave in 60 days, and why?

420 synthetic accounts · 85 testsOnly 16 churn events · no A/B test

Read the case study

Supplier risk shows up only when stock runs low.

Delivery 40% · quality 35% · cost 25%120 synthetic orders · 9 suppliers

Read the case study

An invoice can be paid before anyone checks it.

13 domain tests · a spec error caughtFrontend demo · synthetic · no OCR

Read the case study

Where should the next network point go?

31 API tests · roles, GPS, audit, ROISimulated data · no field test

Read the case study

Smaller projects.

  • SignalOps operations overview with the workspace marked Blocked and KPI tiles marked Blocked or Stale
    Built · tested locally · synthetic data

    SignalOps

    Data-reliability console: a failed check blocks publishing; a replayed batch cannot move a KPI.

    Read the case study
  • The IntelliTix team at the faculty computing exhibition with the app on two screens
    Capstone · Business analyst

    IntelliTix

    IT change-request platform with an AI chatbot; selected for the faculty exhibition, Jan 2026.

  • HeartCheck prototype landing page about starting to look after your heart
    In progress · ML web app

    HeartCheck

    Heart-disease risk from key health indicators. No accuracy or user figures until validated.

  • Notebook comparing classic models with BERT for review sentiment
    NLP · Python

    Netflix review sentiment

    SVM, random forest and logistic regression against BERT on Google Play reviews.

  • Dark dashboard with a world map in an Olympic participation analysis
    EDA · Chart.js, D3.js

    120 years of the Olympics

    Scroll-driven dashboard: participation, gender, national dominance, athlete build.

  • Slide showing the KNN accuracy score and classification report
    ML · Coursework

    Pumpkin seed classification

    KNN separates look-alike seeds: outlier removal, feature selection, 80/20 split.

  • Velocity landing page for advanced activity analytics
    Ongoing · Data viz

    Velocity activity insights

    Strava-style activity analytics in vanilla JavaScript: distance, duration, intensity.

  • GoHealthy screens for planning meals and viewing food history
    Web app · Java Spring

    GoHealthy

    Habit tracker: activity schedule, food log, medication reminders.

  • TasteCraft welcome screen with recipe bookmarks
    Web app · Team project

    TasteCraft

    Recipe search by ingredient or category, moderated comments, shopping lists.

Skills in practice.

  1. Prototype · synthetic dataCapstone

    Requirements & business analysis

    39 gaps · 11 criticalfound in my own spec

    • Gap analysis
    • User roles
    • Approval flows
    Proof

    Reviewed my own retention spec and found 39 gaps, 11 critical, before rebuilding it. Decision Hub turns a free-text brief into seven sourced fields and one approvable quote revision. Business analyst on a capstone selected for the faculty exhibition.

  2. Built · tested locally

    Advanced SQL

    18SQL queries, each hand-checked

    • PostgreSQL
    • Window functions
    • CTEs
    Proof

    18 analytical queries — cohorts, funnel, gaps-and-islands, percentiles, YoY and MoM — each tested against results worked out by hand.

  3. Built · tested locally

    Data modelling & incremental ETL

    53incremental runs match a full refresh

    • Star schema
    • SCD2
    • Python
    • Docker
    Proof

    Star schema with a type-2 history and an accumulating snapshot. 53 incremental runs produce tables identical to one full refresh.

  4. Built · tested locally

    Data quality & KPI reconciliation

    1,848 KPI cells · 0 mismatcheschecked across SQL, BI model and dashboard

    • pytest
    • Data-quality rules
    Proof

    43 automated checks, with sabotage tests that must make them fail. 1,848 KPI cells compared across SQL, the BI model’s logic and the dashboard: 0 mismatches.

  5. Built · tested locally

    Query performance

    1.9× to ~1,150×faster per query after indexing, one laptop

    • EXPLAIN ANALYZE
    • Indexing
    • Partitioning
    Proof

    Before/after query plans on 800,100 synthetic orders: from 1.9× faster for a low-selectivity filter to about 1,150× for a composite index. One laptop, not production latency.

  6. Built · tested locallyPrototype · synthetic data

    KPI & dashboard design

    11KPIs, each defined once

    • DAX (text only)
    • ECharts
    • KPI dictionary
    Proof

    Eleven KPIs defined once, with numerator, denominator, grain and caveats; Power BI-ready views and DAX measures — the DAX has not been run and there is no .pbix. Kinetix shows a supplier score you can take apart: delivery 40%, quality 35%, cost 25%.

  7. Internship

    Spreadsheet analysis on real work data

    Field reportsmatched to live status

    • Excel
    • XLOOKUP
    • INDEX-MATCH
    Proof

    Cleaned and matched field-report data and built a monitoring dashboard that kept each report tied to its real-time status.

  8. Built · tested locally

    Experiment design & statistics

    600simulated experiments, coverage 92.5–97.5%

    • A/B testing
    • SRM
    • Welch & bootstrap CI
    Proof

    Pre-registered protocols and gates that run in order. A deliberately broken test is flagged invalid with no estimate shown; interval coverage stayed within 92.5–97.5% over 600 simulated experiments.

  9. Peer-reviewed · IEEEPrototype · synthetic data

    Machine learning & evaluation

    NDCG@5 0.818 → 0.911offline ranking, not causal

    • Python
    • CatBoost
    • SHAP
    • scikit-learn
    Proof

    CatBoost evaluated with F1, AUROC, AUPRC and Brier score and explained with SHAP, as first author. RetentionOS reports holdout by account and time, a naive baseline and calibration. LARAS ranks offers with NDCG@5 0.818 → 0.911 offline, labelled as not causal.

  10. Prototype · synthetic data

    Applied AI with guardrails

    1 / 36, then 21 / 50held-out cases; gate still fails

    • Claude API (optional)
    • FastAPI
    • Playwright
    Proof

    The model fills seven fields and has no price field; a person approves one exact quote. Karsa keeps a held-out set: its rule-based extractor passes 24 of 24 dev cases but 1 of 36 unseen ones, so its release gate fails. Rewritten and tested on 50 more briefs, it reaches 21 of 50, still below the gate.

Builds run locally, not deployed. Newer ones were coded with AI agents under my direction.

Where I have worked.

  1. Final internship presentation with the Telkomsel team, in front of a screen reading Telkomsel Regular Internship Program
    Final presentation, Telkomsel
    Jul – Sep 2025

    Market & Demand Analysis Intern

    Telkomsel · Fixed Broadband Business · South Jakarta
    • Built a monitoring dashboard tying field reports to real-time status
    • Cleaned and matched report data in Excel (XLOOKUP, INDEX-MATCH)
    • Reviewed fixed-broadband demand with field surveyors
  2. Students working at computers during a practicum session in the Informatics Laboratory
    Practicum session, Informatics Lab
    Feb 2025 – Jul 2026

    Teaching Assistant

    Informatics Laboratory · Telkom University
    • 3 semesters of AI Fundamentals and OOP labs
    • Ran practicums, debugging help, modules and project mentoring
    • OOP tutor in the final semester
    3 lab certificates
  3. 3D render of the Archipelago Pet Collector world: snowy, grassy, cloud, and desert islands linked by bridges
    In-studio render of the game world
    Sep 2026 – now

    Independent Game Developer

    Archipelago Pet Collector · on Roblox
    • 7 regions 284 animations 29 creatures 9 NPCs 38 rigs
    • Fantasy-Nusantara creature collector, built in Blender and Luau
    • Human–AI workflow; 67 audit checks and 39 regression tests
  4. The Informatics for Community team with the supervising lecturer beside the seminar banner
    The team and supervising lecturer
    Mar – Jun 2025

    Team Lead & Speaker

    Informatics for Community · Telkom University
    • Taught AI basics and digital ethics to 22 grade-seven students
    • Led a team of five at a Bandung junior high school
    • Ran the needs analysis, budget and participant feedback review
Members carrying the portable shrine during the Omikoshi parade at Nihon no Matsuri 15
Omikoshi parade, Nihon no Matsuri 15
Beyond class
  • Nihon no Matsuri 15Staff, Nov 2023 – Jun 2024: weekly fundraising sales, maid café timekeeper.
  • GDSC Telkom UniversityMember, 2023 – 2024: challenges and talks, 35+ active participants.
  • Forestry cyber security study groupMember, 2025: security challenges in forest data, new campus lab.
  • Short-form video editor, since 2021TikTok @superrm4x: anime and F1 edits, 127K likes, one at 321K views.
It starts with a question.

How I work.

Before any chart, I write down who will read it and what they must decide.

  1. Ask

    Which decision is this data for?

  2. Clean

    Fix, match, and document the rows first.

  3. Model

    Try the simplest method that answers it, then test it.

  4. Explain

    Show the numbers, the limits, and the next step.

On paper.

Publication · IEEE Xplore

Sensitivity Evaluation of Cardiovascular Risk Profiles for Machine Learning-Based Heart Attack Prediction

First author and presenter · ICICoS 2026, Universitas Diponegoro

CatBoost model; F1, AUROC, AUPRC and Brier score; explained with SHAP.

  • Aug 2026
  • with Untari Novia Wisesty
Education

Telkom University

Bachelor of Informatics, School of Computing.

  • 2022 – 2026
  • Graduation Nov 2026
  • GPA 3.69 / 4.00
  • English test 513
Recognition
  • ICICoS 2026 author and presenterAug 2026IEEE Indonesia Section
  • Capstone picked for the computing exhibitionJan 2026IntelliTix · Telkom University
  • Teaching assistant certificates ×32025 – 2026AI Fundamentals, OOP, AI & OOP
  • Nihon no Matsuri 15 staff certificates ×2May – Jun 2024Fundraising, café timekeeper
Certificates

A few
answers.

Contact
What roles are you looking for?

Business analyst, data analyst and applied AI roles or internships, in Bandung, Jakarta, or remote. I am at my best between understanding the business problem and building the system that makes the analysis usable.

Are these projects real?

They are working prototypes with automated tests, on synthetic or simulated data, and every case study says so; no client or business result is claimed. The newer builds were implemented with AI coding agents under my direction, and each case study states my role.

Project details