Nauval Zulfikar

Problem

The Aston PhD project description (p196342) calls for a Digital Twin framework that integrates technical IPFT operations with behavioural responsiveness under disruption. No publicly available reference architecture demonstrates this hybrid simulation pattern at scale.

Outcome

• —Smoke-tested: 45% acceptance rate over 20 ticks · 62% terminal utilisation
• —12 files, 1045 lines, publicly available on GitHub
• —Referenced as the architectural prototype of Phase 1 in the PhD research statement

Problem

The Aston PhD posting calls for 'use of Large Language Models (LLMs) to generate adaptive user profiles and decision rules' — a methodology with no public reference implementation. Existing crowd-shipping decision models rely on fixed-utility discrete choice, missing the heterogeneity LLMs can capture.

Outcome

• —13 files, 1449 lines, publicly available on GitHub
• —Demo notebook: 5 walkthrough examples + aggregate threshold distribution
• —Integration hook into Mini-Digital-Twin agent-based simulation layer

Problem

The Aston PhD posting names 'optimisation' as an essential requirement. The Archetti, Savelsbergh & Speranza (2016) VRPOD is the canonical formal model for crowd-shipping, but the reference implementation is not publicly available.

Outcome

• —13 files, 793 lines, publicly available on GitHub
• —Toy instance result: baseline cost 63.84 → with crowd-shippers 39.29 = 38.5% savings
• —Phase B extension plan: explicit MTZ subtour-elimination, scale to 20-50 customers

Problem

The crowd-shipping literature posits non-linear reward elasticity (Punel & Stathopoulos 2017) and supply-side matching effects (Le et al. 2019), but most agent-based studies focus on a single parameter axis. A research-grade ABM should systematically map the joint reward × supply response surface.

Outcome

• —12 files, 1075 lines, publicly available on GitHub
• —Supply effect: acceptance rises from 45.34% (n_pax=20) to 79.80% (n_pax=100) — 34pp spread
• —High-high corner (reward ≥ £40, n_pax ≥ 80): 78% acceptance ceiling

Problem

Indonesia's school nutrition programme serves millions of meals every day, yet kitchen operators run on spreadsheets — menu planning, ingredient variance, delivery routing, and coordination across multiple kitchens all break the moment one operator runs more than a single kitchen.

Outcome

• —Live in production for the pilot kitchen (SPPG Paseh)
• —Variance reporting replaces manual reconciliation between planned menus and actual deliveries
• —Foundation for the GiziGuard on-site AI nutrition layer (next project)

Problem

Kitchens under the school nutrition programme need real-time nutrition validation, but schools have unreliable internet and strict expectations that data stays on-premises. Cloud-hosted AI is neither reliable nor appropriate here.

Outcome

• —Submission targeted for the Kaggle Gemma 4 Good Hackathon (deadline 2026-05-18)
• —Reusable on-site AI layer applicable to other government nutrition programmes

Problem

Bandung's permits department had no single view over regional data scattered across the national permits portal, spreadsheets, and manual department reports — covering building permits, small-business records, tourism, and spatial plans. Leadership read static weekly reports that were already stale by the time they arrived.

Outcome

• —Live at sibedaspbg.cloud, in daily use by department staff
• —Leadership self-serves answers from the data instead of waiting for weekly reports
• —Automatic fee calculation replaces a manual process that used to produce errors

Problem

Bandung's residential property market is scattered across multiple listing sites, brokers, and direct-from-owner channels. Buyers and agents waste hours collating listings, and direct-owner inventory buried under broker stock gets missed entirely.

Outcome

• —Live admin running with active aggregation across the Bandung area
• —AI chatbot acts as a buyer-side property advisor over the aggregated catalogue