Joe Eppel

Problem. Bitcoin holders face a critical gap: without a plan, their assets become permanently inaccessible to loved ones. Xapo Bank identified this as a core unmet need for long-term holders.

Approach. Led end-to-end product definition for a secure beneficiary designation feature. Worked with legal, compliance, and security teams to leverage our existing MPC-based custody solution that combined Swiss vault-grade custody with a simple, trustworthy UX flow for the account holder to easily nominate and identify their intended beneficiaries of their multi-currency wealth.

Outcome. Launched November 2024. Shortlisted as a Diamond Finalist at Money20/20 Europe 2025 — recognition from peers and press as a genuinely novel product in the space.

Problem. Xapo Bank's onboarding funnel was losing the majority of applicants before they completed KYC — a critical step in a regulated product. The drop-off points were unclear and the team lacked the instrumentation to run structured experiments.

Approach. Implemented Amplitude across the funnel to get behavioural visibility, then built an experimentation cadence of ~30 tests per year. Identified friction in the ID-verification step and redesigned the flow based on data. Made experimentation a company-wide discipline, not just a product team habit.

Outcome. Doubled overall funnel conversion rate and increased ID-verification starts by 50%. The work was featured as an Amplitude case study.

“Now, our whole company thinks about metrics, not just the product team.”— Joe Eppel, in the Amplitude case study

Problem. Cape Town has dozens of beaches with wildly different conditions on any given day — wind, temperature, water quality — and no easy way to tell which one is worth the trip. I also wanted a real end-to-end test of building a product solo with AI.

Approach. Built a scoring system that combines air and sea temperature with wind speed, pulling live data from Open-Meteo and beach metadata — amenities, water quality, lifeguards — from the City of Cape Town, with filtering across 27 beaches. Designed, built, and shipped it end-to-end working solo with Claude.

Outcome. Live at nowcast.capetown, giving Cape Town beachgoers a real-time score for every beach in the city.

Problem. Our school sends newsletters packed with dates — closures, deadlines, sports days, fee cutoffs — buried in paragraphs of text (and sometimes a PDF for good measure). I kept missing things.

Approach. A scheduled pipeline polls Gmail for new newsletters, pulls text from the body and any PDF attachments, and hands it to Claude with a structured extraction prompt anchored to the email's send date rather than the processing date — so relative dates like "this Friday" don't drift when a backlog of unread mail builds up. Ambiguous timing ("later this term") is deliberately left out of the calendar rather than guessed. Extracted events are deduplicated and written straight to our family Google Calendar, alongside a summary email and a weekly two-week look-ahead digest.

Outcome. It's saved us from at least a few "wait, school's closed today?!" moments — and gave me a real, low-stakes place to get careful about prompt-injection handling and date-logic bugs, since the input is coming straight from an inbox I don't fully control.

Problem. Bank statements pile up as PDFs and CSVs with no consistent categorization, and the cloud finance apps that could make sense of them require handing over transaction data to a third party. I wanted month-by-month visibility into spending without any of it leaving my machine.

Approach. Joumoney imports bank statements through bank-specific parsers — Investec, Discovery, and more — with generic CSV/PDF fallbacks for the rest, then runs each transaction through a hybrid categorizer: a small set of seed rules catches obvious cases like payroll and brokerage transfers, backed by a TF-IDF + logistic regression classifier trained on my own correction history. Predictions are confidence-gated, so only high-confidence categorizations auto-apply — everything else routes to a quick review queue rather than silently guessing wrong. A second, scoped classifier assigns optional subcategories (Transport → Uber) constrained to the parent category, and hash-based dedup keeps overlapping statement re-uploads from creating duplicate transactions.

Outcome. Every fix I make during review becomes a training example, so the categorizer gets more accurate the more I use it — and a monthly dashboard shows where the money actually went, trended over time, entirely on-device. Backed by 26+ pytest modules covering parsing, categorization, and storage.