Propose before writing
Invoice drafts enter a human approval workflow before an authorized write can occur.
THE TRADE-OFF
An extra approval step sacrifices full automation in exchange for explicit control.
AI PRODUCT DEVELOPMENT · FINTECHLocal prototype
FinPilot.
Making financial automation more reliable through human oversight.
Exploring financial AI that proposes actions and earns permission to execute.
Explore the repository (opens in a new tab)How much autonomy should a financial workflow have?
Financial operations combine repetitive retrieval with consequential decisions. FinPilot explores assistance for customer lookup, invoice search, policy retrieval, duplicate detection and invoice proposals—while keeping authorized writes behind explicit human approval and deterministic controls.
- MY CONTRIBUTION
- Product requirements, workflow design & technical implementation
- WHO IT’S FOR
- Finance operations teams handling invoice and policy workflows
- PROJECT STAGE
- Local prototype
- EVIDENCE AT A GLANCE
- 52 passing tests · 150 offline scenarios
A useful assistant needs a trustworthy boundary.
The project frames repetitive financial operations as an opportunity for assistance. But incorrect amounts, duplicate invoices and unauthorized writes can outweigh the convenience of a faster answer.
This is a technical product prototype using synthetic records, not a customer-validated production deployment. Its central design question is how to separate helpful retrieval and proposals from actions that change financial records.
52
Passing automated tests
Independently rerun on November 10, 2026.
150
Offline evaluation scenarios
Synthetic and deterministic; not live LLM accuracy.
1
Explicit approval boundary
Proposals are separated from authorized writes.
THE PRODUCT INSIGHT
The AI can propose an action. The system must independently decide whether that action is authorized and valid.
The choices that shaped the product.
Make controls deterministic
Validate financial inputs and approval state; enforce idempotency and duplicate checks outside the language model.
THE TRADE-OFF
Rules require maintenance, but critical behavior becomes testable and repeatable.
Keep AI integration optional
A local deterministic path works without a paid API; optional LangGraph and OpenAI support agent exploration.
THE TRADE-OFF
Offline scenarios verify safeguards, not the accuracy or resilience of a live language model.
A decision you can walk through.
Explore the product thinking through a small, interactive illustration. The example content below is fictional and is not a screenshot or a live product integration.
INTERACTIVE ILLUSTRATION Synthetic data · Workflow illustration
THE HUMAN DECISION POINTA proposal is not
A proposal is not
permission to write.
Try the approval boundary. Approving creates one illustrative local draft; rejecting creates none. This interaction does not call FinPilot or move money.
Amount validated Duplicate check passed Human review required
SAMPLE INVOICE PROPOSALReview needed
- Customer
- Example Studio Synthetic
- Amount
- $1,200.00 USD
- Reference
- DEMO-001
0 drafts created. Awaiting a human decision.
ILLUSTRATIVE AUDIT TRAIL proposal.created validation.passedawaiting.review
A boundary at every consequential step.
Request & retrieval
FastAPI exposes customer, invoice and policy workflows.
Optional agent
LangGraph can orchestrate tools with an optional OpenAI provider.
Proposal & validation
Deterministic controls inspect the draft and prevent invalid operations.
Human approval
Approval state is checked independently before committing a write.
Local write & audit
SQLite and synthetic records support idempotency and audit trails.
Python, FastAPI and local SQLite are source-inspected. Optional live OpenAI/LangGraph execution is not independently verified. No real banking, payment, QuickBooks, or production financial integration is claimed.
What the evidence actually supports.
Controls tested in a local environment
All 52 automated tests passed when independently rerun for this portfolio on November 10, 2026. They cover validation, approval, idempotency and resilience. These local tests do not establish production reliability.
Evaluation has a specific boundary
All 150 synthetic offline scenarios also passed on rerun. They exercise deterministic behavior. They are not evidence of live LLM accuracy, real-world financial correctness, adoption, or production readiness.
What I’m taking forward.
A trustworthy workflow is more than an impressive AI demonstration. The proposal, approval and execution boundaries are part of the product experience.
The next useful step
Evaluate live agent behavior separately with adversarial requests, ambiguous records and tool failures. Expand audit review and operator usability before considering real financial data.
Sources & evidence notes
Reviewed November 10, 2026. Repository documentation is attributed to its author; implementation claims are distinguished from research findings and planned capabilities.
Repository README & implementation scope (opens in a new tab)Product requirements (opens in a new tab)Technical architecture (opens in a new tab)Evaluation methodology & limits (opens in a new tab)Download the content verification record