Constantine Samarikas

Vendr

Free collection, pricing and trading app for Pokémon and One Piece card collectors and shops in Europe, priced from Cardmarket in euros. Built solo, from first idea to a working app on iOS, Android and the web.

• 38,329Pokémon and One Piece cards, priced per language
• 5.8Mprice records since November 2024, about 75,000 a day
• 5price feeds across Cardmarket and US markets, merged nightly

• One pricing rule, implemented in the app, the pipeline and the database, kept identical by golden test cases that run against all three on every push.
• Product links are verified against Cardmarket's catalogue and locked in the database; a freshness alarm fails the nightly run when any source goes stale or its volume drifts more than 50% from normal.
• Card scanning from a photo: native card detection in Swift and Kotlin, an on-device image model and vector search across the whole catalogue.
• Developed with a reviewed AI-agent workflow: separate agents write and review each change, and nothing ships without approval.

• TypeScript
• React Native
• PostgreSQL
• Supabase
• pgvector
• PyTorch
• Swift
• Kotlin
• GitHub Actions

Operationalising Regulatory Red Flags for Bitcoin Anti-Money Laundering Detection

BSc dissertation, The University of Manchester

Turns the six FATF virtual-asset red flags into 35 measurable features and detects illicit activity without large hand-labelled datasets, using rule-based pseudo-labelling, clustering and LightGBM on data from the Cardiff University Bitcoin Database.

• 0.96macro F1 on a balanced evaluation set
• 8.3Mtransactions, 24 quarterly snapshots, 2019–2024
• 58.6% → 7.5%mixing-related flags, Q3 2023 to Q3 2024

• 0.96 macro F1 on a balanced evaluation set of 225,722 transactions (the 149,128 confirmed labels plus a sample of unmatched transactions presumed licit); 0.93 on the confirmed labels alone. About 12 percentage points above the strongest combination of the rules themselves.
• Clustering run without the labels isolated 855,945 transactions holding 44,509 confirmed illicit and only 221 confirmed licit: 99.5% purity, and about 39% of all confirmed illicit activity in the data.
• Matched or exceeded published results on the Elliptic++ benchmark: the same precision (0.99), with recall up from 0.73 to 0.97.
• Rule firing tracked enforcement: mixing-related flags peaked about a year after the Tornado Cash sanctions, then fell as Samourai and Wasabi faced enforcement.