I build entire production
systems end-to-end.

I joined a UK charity where the technology was in fragments. The website had been built by an external agency. Every donate button on the site still routed through their hosted checkout — a system the charity didn't own or control. The mobile donation experience took over 30 seconds to navigate, with double form fills and a clunky UI. Donations were scattered across three separate platforms — the agency's checkout, Enthuse, and LaunchGood — each with different payout cycles (2 days vs weekly vs sometimes two months), different fee structures, and inconsistent Gift Aid handling. The CRM integration was broken: donation data wasn't flowing into the system properly. Gift Aid reconciliation was largely manual.

Over 18 months as the sole developer, I systematically replaced every external dependency — 14 repositories in total. Built a new checkout from scratch — multi-step with progressive disclosure, Stripe PaymentIntents for one-off, SetupIntents with off-session charging for recurring (the system only allows amount increases on active plans — a deliberate constraint). Apple Pay and Google Pay. Three separate donation builder components with Zod validation. Built the P2P fundraising engine — individual fundraiser pages, team pages, leaderboards, URL-based attribution, all backed by a Prisma schema with 15 interconnected models. Built a config-driven campaign landing page system with a 30-field campaign interface — slug, amounts, impact messaging, widget type, social proof, custom HTML support — so the fundraising team could launch Ramadan appeals without waiting for a developer.

Then I built the data layer. Three separate sync pipelines for each donation platform: Enthuse required Playwright browser automation with a four-layer auth chain — SSO login triggers a magic-link email, Gmail OAuth reads it, Playwright clicks it, then optional TOTP via speakeasy. LaunchGood required Playwright to navigate the export UI, check an acknowledgment checkbox, and capture the CSV download stream. Both use session persistence via storageState.json, safeClick() retries, and screenshot dumps on failure. A third pipeline syncs the charity's own MySQL database into PostgreSQL — five data streams (one-time donations, scheduled recurring, supporters, appeals, legacy donations), each joining 10+ source tables, landing in raw tables then transforming into a canonical model. That canonical model feeds N3O CRM queue payloads and a donor messaging system built on n8n + Chatwoot + WAHA with a database-backed send queue. Terminated the agency retainer. Migrated hosting.

The same charity needed to reach potential partner organisations across the UK charity sector — hundreds of thousands of registered charities in the Charity Commission dataset. The fundraising team had no tooling for this. They were manually searching, manually qualifying, manually writing emails. It didn't scale and the outreach was generic.

I built a pipeline that ingests raw Charity Commission data into PostgreSQL, then uses OpenAI to translate natural language queries ("large education charities with income over £1M operating nationally") into SQL filter logic via a custom segment engine. Once leads are qualified, OpenAI generates personalised outreach — emails, talking points — based on each charity's actual profile, income band, and operational focus. Not templated mail-merge. Actually personalised against real data. The system enriches contacts through Apify to find the CEO, Director, or Head of Fundraising. The whole thing runs from a CLI with deterministic Python scripts underneath.

Charities using third-party pledge platforms were losing money to fees, waiting weeks for payouts, and had no control over the donor experience. The existing options — Enthuse, LaunchGood — were slow to pay out, inconsistent with Gift Aid handling, and charged admin fees the charity couldn't avoid. I built an alternative from scratch.

Full pledge checkout with Stripe integration — commit now, pay later, with three payment rails (bank transfer, GoCardless direct debit, card). Pledge lifecycle tracks through five statuses: new → initiated → paid → overdue → cancelled. Built-in bank statement reconciliation with configurable column mapping. But the real story is the AI "Steward": an embedded assistant that helps fundraisers manage their campaigns. It uses OpenAI with Gemini fallback, structured function calling with 13 tools that return rich visual cards — stats grids, pledge tables, campaign detail, WhatsApp status, nudge drafts. Every response is grounded in real pledge data.

I stress tested it across 31 scenarios in 10 categories — both bash and Python harnesses. Dashboard overview, pledge lookups, campaign queries, setup status, action requests, navigation, edge cases (gibberish input, dangerous requests like "delete all our data"), brand compliance (regex checks for forbidden terms: "AI assistant", "language model", "ChatGPT", "artificial intelligence"), persona scenarios (event lead, treasurer asking about Gift Aid, attribution queries, pledge amendments, trustee report prep), and tone validation (gives specific advice, responds with empathy not shame). 31 of 31 pass.