All projects

Three-layer Digital Twin prototype for behaviourally-responsive crowd-shipping: FastAPI mock-IoT data layer, mesa agent-based + simpy discrete-event simulation, Streamlit dashboard with interactive disruption injection. Built for Aston PhD application p196342.

End-to-end system that generates literature-grounded crowd-shipper decision rules from traveller profiles, using RAG over a behavioural literature knowledge base + gpt-4o-mini. Direct implementation of the methodology named verbatim in the Aston PhD project description.

Mixed-Integer Linear Programming formulation of the Vehicle Routing Problem with Occasional Drivers (Archetti et al. 2016), solved with scipy.optimize.milp (HiGHS). On a toy 8-customer / 2-truck / 3-shipper instance, crowd-shipping reduces total cost by 38.5%.

Agent-based simulation of crowd-shipper acceptance with a systematic parameter sweep characterising the reward-supply response surface. 450 simulations: 9 reward levels × 5 supply densities × 10 replications. Auto-generated findings cite Punel & Stathopoulos (2017) and Le et al. (2019).

Replaced spreadsheet juggling for the school nutrition programme — one system handles menu planning, delivery routing, and variance reports across multiple kitchens. Each kitchen's data stays cleanly separate, verified by 96 automated checks.

Early-warning system for rolling-element bearings, built with PT Bara Sakti Ziwa Abadi. It reads raw vibration, compares each bearing against its own healthy baseline, and names the fault from bearing geometry with no training data. Validated on three public run-to-failure datasets and published as an interactive bilingual dashboard with a procedural 3D bearing simulator.

Replaced static weekly reports for Bandung's permits department — leadership now queries KPIs in plain language. Permit fees calculate automatically from rules the policy team can update themselves, no developer needed.

Pulls real-estate listings from across the Bandung market into one place, drafts Indonesian-language marketing copy automatically, and surfaces direct-from-owner properties that brokers usually bury under their own stock.

LLM-powered job application platform — evolution of the Nov 2024 CV-JD suitability research into a full product.

Production point-of-sale platform powering live retail operations — multi-branch admin, offline-first cashier app that keeps selling when WiFi drops, separate per-tenant data.

Personal quant finance workbench on MT5 broker data, focused on XAUUSD.

AI-powered language learning platform built on Spaced Repetition, Active Recall, and Comprehensible Input.

Product catalogue analysis platform — rebuilt from one-off scripts into proper software with a clean admin UI.

Internal accounting platform with hierarchical Chart of Accounts and a bank-statement reconciliation engine.

Affiliate-first commerce automation system targeting <2 hours manual intervention per day at scale.

Web tool to author Claude Code agents directly into Google-Drive-synced project folders, with image context.

Customer-facing app for managing customer logos and brand assets, backed by an AI tagging assistant.

Landing site and brand hub for the Aureonforge project cluster, fronting DPMBG, Sibedas, QF, and more on a single server.

Internal admin platform for managing dozens of concurrent infrastructure projects at Bandung's Public Works & Spatial Planning Department — budgets, vendors, milestones, sign-offs, all in one place.

Consolidated P&L for a multi-entity, multi-currency group (9 entities in USD, CAD, EUR and AUD, plus eliminations). Splits the budget miss into exchange-rate effect and operational variance, and shows what minority shareholders own.

B2B pipeline report fed from a CRM API: an incremental extractor, a DuckDB warehouse that keeps every version of every deal, and a Power BI model with incremental refresh. Shows how much of the open pipeline is real.

Inventory and purchasing report for a 79-store wine and spirits chain. A restock engine computes safety stock, reorder point and order quantity per store × item, and recalculates in DAX when the service level or review period changes.

Sales-against-quota report for a bike wholesaler with 17 reps in three branches. Power Query cleans each branch's messy Excel format into one sales table and logs every row it sets aside, with the reason.

Operations report for four engineering teams built from their real public Jira tickets (Apache Hive, Cassandra, HBase, Artemis). Maps each team's own workflow onto shared stages and replays every number from the full change history.

Credit-risk report on 2.26 million LendingClub consumer loans (2007–2018): what was lent, what came back and what was lost, with a probability-of-default model and expected loss on the live book.

Sales, customer, delivery and seller report for Olist, a Brazilian marketplace: about 100k orders from 2016 to 2018 across 9 linked tables, turned into a star schema and a five-page Power BI report.

Every Grand Prix from 1950 to 2024: 1,125 races, 861 drivers, 211 teams and 77 circuits, from 14 raw files to a star schema and a six-page Power BI report, with a driver ranking that removes the car's advantage.

Proposal and ML feasibility analysis for automating dam and irrigation gate control.

Comparative performance analysis of multiple ML models for footwear sales forecasting.