Deep dive · Finamaze · shipped
Finamaze: investor onboarding, KYC and reconciliation
The investor onboarding flow with document and visual identity checks, and reconciliation that keeps the platform's ledger in step with the bank.
Problem
Finamaze runs a robo-advisor that builds AI-tailored portfolios for each investor’s risk profile. Before anyone can invest, the platform has to know who they are, and once money moves, its own ledger has to agree with the bank’s records to the cent. Both jobs used to depend on people checking documents and spreadsheets by hand.
Constraints
- Regulated onboarding: on an asset-management platform, identity checks are a precondition for investing, and onboarding has to follow the regulator’s policies.
- Two kinds of evidence: the document itself and the person presenting it both need verifying.
- Money must reconcile: any difference between the platform’s records and the bank’s has to surface quickly, not at month end.
- Privacy: identity documents are among the most sensitive data a fintech holds.
Architecture
Diagram source (Mermaid)
flowchart TB
signup[Sign-up · web and mobile] --> docs[Documents · selfie]
docs --> dv[Document verification]
docs --> vv[Visual verification]
dv & vv --> ok[Approved]
dv & vv --> review[Review queue · unclear cases]
ok --> account[Investing account]
account --> recon[Reconciliation · ledger vs bank records]

Key decisions
- Document and visual verification as separate checks. A valid document says nothing about who is holding it. Verifying the document and matching the person to it are separate steps with separate outcomes, so a failure says which one failed.
- A review queue for unclear cases instead of automatic rejection. Rejecting every borderline case loses real investors; approving them loses compliance. Unclear results go to a person with the evidence attached.
- Reconciliation as a continuous job, not a month-end report. The broker’s data is the source of truth: the platform’s ledger is matched against it on a schedule, so every difference is found and traced while it is still small and explainable.
Results
- Manual review time per file
- −60%