Dylan Mérigaud, freelance AI engineer for fintech

the trace

Selected work

Three public reference systems. Live where a browser can show it, every repo open.

01 · match / ledgerloop

ledgerloop

An agent reads a company's HRIS and derives the whole approval workflow. Then invoices route through it.

A vendor PDF is extracted by a vision model, matched 2/3-way against open purchase orders captured from a live QuickBooks org, and routed through an approval DAG derived from real BambooHR data. Nobody draws a workflow canvas: the agent derives it from the org chart and you maintain it in plain language.

AI sits only where it earns its keep: reading messy documents, mapping org titles to signing authority, investigating flagged exceptions. The money path is deterministic, unit-tested code, and nothing posts before a human approves. Every run lands as an append-only audit row, replayable with zero tokens. Swap the QuickBooks adapter for NetSuite and nothing downstream changes.

Live demo

Mastra
Next.js
TypeScript strict
Zod
Drizzle
QuickBooks
BambooHR

02 · route / approvals-ui

approvals-ui

The approval workflow screen, as shadcn components.

Quorum gates, amount thresholds, a policy lint that knows what segregation of duties means, and plain-language editing where a human reviews the diff before anything lands. One command and the code lands in your project, yours to edit. Built on react-flow-auto-layout, my published npm package that lays out React Flow graphs the way dagre should.

Live playground

shadcn registry
React Flow
react-flow-auto-layout

fig. 02 · plain-language edit · approvals-ui

03 · audit / fintech-roast

fintech-roast

An agent that roasts the code that touches money.

A rulebook of 41 researched rules across 10 domains, from ledger integrity and rounding to FX, tax, webhooks, and time, applied per-language to TypeScript, Python, and Java, with every finding adversarially verified before it is reported. Read-only: it never edits your code. Run it on your own repo in two commands.

Claude Code plugin
41 rules
read-only

fig. 03 · verified findings · fintech-roast

also in the trace

AI Invoice Parser, a schema-validated extraction demo with an eval harness across 9 messy real-world formats. Live demo · GitHub

04 · trace / nine years of runs

Pivot · procurement fintech, Paris

Shipped the PO approval flow and the NetSuite integration. Cut client onboarding time ~90% by automating the approval setup.

Neige · founder

AI content-generation pipeline. Empty repo to production, solo.

Runtime · 9 years

Full-stack, exclusively in startups and scale-ups.

05 · policy / no llm on amounts

An LLM never decides a payment amount.

Deterministic where it must be, agentic only where the trajectory is genuinely open-ended. AI reads the documents, derives the workflows, and investigates the exceptions. Code moves the money.

Next.js
TypeScript
Node
Python / FastAPI
Postgres
Supabase
Mastra
OpenAI & Anthropic APIs

06 · scope / how we start

Work with me

Start with a paid pilot.: One scoped problem, a fixed one to three week sprint, a deliverable you run in your own repo. Low risk for you, fast proof for both of us. It grows into a larger build or it doesn't, and either way the work is yours to keep.
A bridge while you hire.: A senior hire takes months to land. I ship the AI piece that can't wait now, on a defined scope, while you run the search. The contract ends whenever you want it to. Open to a permanent role later if it clicks on both sides.
Available now.: Remote from Mexico City (GMT-6, full overlap with US hours), on-site in NY or SF for key milestones, Paris too. Fixed-scope pilots and two to four week builds.

Pilots and builds are fixed-price, hourly for open-ended work, rates on the call.

07 · approve / let's talk

Building AI into a fintech product? Let's talk.

dylan@merigaud.com Book 30 min

Remote from Mexico City, working US hours (GMT-6). On-site in NY/SF for key milestones. Paris available too.

trace complete · human approved