What I build

Tier-1 payments provider

Surfaces related past incidents and writes the queries on demand. Average triage went from roughly an hour to roughly twenty minutes. It investigates and recommends; a human decides and acts.

• Retrieval over incident history
• Query generation on demand
• Human in the loop by design

Inside the perimeter

RAG, structured extraction, agentic tool use and on-device LoRA fine-tuning, all running on my own hardware. No data leaves and no invoice arrives from an API.

• RTX 3060, 12GB
• Qwen 2.5 Coder 14B Q4
• Context is the real ceiling

Built in public

Three languages, a database behind a fully static site, twelve dollars a month on GCP. Every architectural decision becomes a video.

• NestJS + Prisma + Postgres
• React prerendered at build
• Cloud Run scaling to zero

Multi-tenant B2B platform

Property management for buildings, sold to the operator rather than to the resident. Four backend services around a shared library, each a modular monolith, with access control resolved on two deliberate axes: what an endpoint allows, and what a plan actually includes.

• Java 25, Spring Boot
• Postgres and Redis on GCP
• Tenant is the payer, not the building

Single-tenant B2C, tarot

A tarot platform that puts the game, the learning and the community in one product. Architecturally it is Condo turned inside out: the user buys, so entitlements resolve per person instead of per organisation. Building both at once is what keeps the shared foundation honest, since anything that only works for one of them was never really shared.

• Shares the platform library
• User-centric entitlements
• Single tenant at the database level