finance professional → AI systems builder
I build AI tooling for finance work that has to be right — extraction pipelines, verification layers, and the internal platform ~30 people use daily at a Dublin accounting practice. Background in global-mobility tax (PwC Japan / Vialto Partners). Trilingual: English, Japanese, Filipino.
Featured
An extraction + verification agent for financial filings, built to be measured, not believed.
Reads public SEC 10-Ks, extracts key figures with an LLM, then verifies them with deterministic code — balance-sheet arithmetic, cross-statement consistency, roll-forward logic — under a full evaluation harness. Extraction scored against each filing's own XBRL as oracle: 100% recall on all 11 target fields across 19 large-cap filings, 22/22 exact match on the development filing. Synthetic error injection: 61 injected cases, 100% caught, 0 missed. And the finding that matters — false-alarm rates measured per check on clean data, including one check that fires on 81% of clean filings, with the full analysis of why and what fixes it.
Tools built solo inside an accounting practice — capability and results shown; client data stays where it belongs.
ProblemForensic casework starts with scanned bank statements — historically transcribed by hand, line by line.
BuiltAn AI extraction pipeline that converts scanned statement PDFs into structured, analysis-ready transaction data.
Result3,018 transactions across 106 scanned pages in ~30 minutes for under €0.50 — roughly 70× faster than the manual process — validated by a partner on a live forensic case.
ProblemDivorce and litigation casework requires categorising thousands of transactions consistently and defensibly.
BuiltA categorisation engine — 746 keywords across 21 categories — producing structured workbooks for forensic review.
ResultIn use on live matrimonial and litigation cases; consistent categorisation at a scale manual review can't match.
ProblemVAT return preparation is repetitive, deadline-driven, and error-intolerant.
BuiltAn automation pipeline with confidence scoring and duplicate detection, so human review concentrates on the entries that need judgment.
ResultIn production across the firm's client base.
ProblemA 30-person practice adopting AI with client data in play — no policy, no visibility, no way to judge what's worth paying for.
BuiltThe firm's AI operating framework: usage policy, per-tool GDPR/PII assessment, a usage-tracking dashboard, and an ROI framework for the toolstack.
ResultFirm-wide rollout with guardrails people actually follow — built by the person shipping the tools.
ProblemClient records, deadlines, and work status lived in spreadsheets and memory.
BuiltA self-hosted practice platform with a 12-column sortable dashboard, activity logging, and bulk operations. Designed, built, deployed, and maintained solo.
Result329 client records, ~30 daily users, running in production since March 2026.
Open Brain — my own MCP memory server: semantic search over captured thoughts, scheduled consolidation, browse UI, voice capture. Self-hosted; used daily as external memory across AI sessions.
Voice-note transcriber — local Whisper + Claude pipeline: folder watcher, transcription, action extraction, desktop notifications.
Newsletter automation — Claude-drafted, partner-approved by email, auto-published to Teams and broadcast to 28 staff (Resend + Power Automate).
Earlier public builds — Forensic Transaction Analyzer · Invoice Processing Agent · Tax Q&A Bot — code on GitHub
Global-mobility tax at PwC Japan / Vialto Partners, then Dublin, then the firm's AI and automation builder. My real skill is AI-assisted build and delivery — shipping a working tool fast, then hardening it with the people who use it. I care most about the trust layer: evaluation, verification, and knowing exactly where a system fails.
BBA Business Economics, Tokyo International University. English · 日本語 · Filipino.