In a regression, the residual is what's left after the model has said everything it can. That's the part I like: take a headline number, find the public data behind it, and see what the average leaves out.

I'm a final-year B.Sc. (Hons.) Statistics student at UiTM, graduating in February 2027, and currently a Business Analyst intern at Alto Soft Sdn Bhd, where I prepare UAT and write user manuals and SOPs for a multi-module government system. Most of what's here is built on Malaysian open data—OpenDOSM, the Ministry of Higher Education, MetMalaysia, and the federal Budget documents—alongside production-style pricing, survival, and streaming pipelines.

Budget 2027 announced a RM2,500 monthly wage floor for graduates. Half of 2024's first-degree graduates were already at or below it (49.7%), but that half is not spread evenly across disciplines:

The numbers behind the chart live in data/, and the full microsimulation is in fresh-grad-salary-model.

1. fresh-grad-salary-model — How many graduates does a RM2,500 wage floor reach, and what does it cost?

   49.7% of 2024 first-degree graduates earned RM2,500 or less. In simulation, the graduate wage bill rises 7.1% to 14.4%, depending on who is covered and how spillover behaves. (Python, statsmodels, Excel)

2. pricecatcher-retail-pricing & pricecatcher-anomaly — Store-format pricing and weekly profiteering leads across 20M+ prices

   A 16-staple basket is 14% cheaper in hypermarkets (RM110.96) than wet markets (RM126.58) under two-way fixed effects, though within-format shop dispersion dominates. To shortlist weekly inspections across 16.6M prices (Jan–Oct 2026) without profiteering labels, planted +20% markups served as the benchmark: a simple peer-grouped robust z-score caught 21% of held spikes at top-20/district capacity vs 9–11% for Isolation Forest and ECOD. (DuckDB, Python, scikit-learn, Power BI)

3. hr-attrition-decision — Who quits a job, and is it worth paying to keep them?

   Real survey data on 1,116 employees shows staff on partly undeclared ("grey") pay stay a median 26 months vs 54. A Cox model ranks risk modestly (C-index 0.60), and back-testing the retention cost rule beats "treat nobody" or "treat everybody" in just 4 of 45 cost scenarios—meaning treatment-effect estimation matters more than the model. (Python, lifelines, Power BI)

4. ev-conformal — How honest is a published paper's "90% range"?

   Re-examining arXiv:2610.13588 on EV charging uncertainty: an honest calibrated method would pass only ~53% of the paper's scorecard, so its recommended method wins by being wider. Three code discrepancies were identified and reported to the authors. (Python, conformal prediction)

• malaysia-inclusive-economy-blueprint · live — Interactive fiscal policy simulator.

• perasaan-malaysia — Scheduled pipeline for MetMalaysia's 7-day forecasts.

• esg-multiagent-auditor — Event-driven Scope 1–3 GHG auditing architecture.

• free-worker — MCP worker for Claude Code that routes tasks across free AI models and falls back when one hits its quota.

• threads-mcp — Read-only Threads API connector for topic analysis.

• hf-agents-course-gaia — Hugging Face Agents Course final agent (85% on GAIA Level-1 questions).

• wealthwatch · live — Personal finance tracker.

                        Model summary: luqman
======================================================================
Dep. variable:   answers you can check
Method:          least squares
Sample:          Oct 2023 - Feb 2027, B.Sc. (Hons.) Statistics, UiTM
Current:         Business Analyst intern, Alto Soft Sdn Bhd (Oct 2026 -)
----------------------------------------------------------------------
Notes:
[1] Every number comes from a linked repository or from GitHub.
[2] Simulated data is labelled as simulated.
[3] The header and chart are drawn by scripts/build_assets.py;
    the activity image is refreshed daily by lowlighter/metrics.
======================================================================