The problem was not another chatbot
Financial review depends on documents, rules, exceptions, and a record of how a conclusion was reached. I built FinSight to explore an AI workflow that could extract information, detect inconsistencies, perform KYC and AML checks, and produce a result that a reviewer could trace. The core requirement was accountability. A polished answer without evidence would not be useful in a compliance setting.
Separating extraction from reasoning
The workflow begins with LandingAI ADE extraction so document fields become structured records rather than ungrounded text. FastAPI coordinates the backend while React presents the review. I kept extraction, validation, reasoning, and report generation as separate stages. This makes it possible to see whether an error came from a misread field, a failed rule, or an unsupported inference.
Building checks that can be audited
The system compares identities, dates, addresses, and financial details across documents, then flags conflicts for review. LLM-based validation helps explain relationships, but deterministic rules remain visible where the requirement is explicit. Every result includes supporting values and traceability back to the source material. Sensitive actions remain recommendations for a human reviewer rather than automatic decisions.
What I learned
FinSight taught me that responsible AI architecture often looks less magical because it exposes its steps. That is a strength. In high-impact domains, the product should make uncertainty and disagreement easy to find. My next iteration would deepen access controls, retention policies, evaluation datasets, and reviewer feedback loops. Accuracy matters, but governance determines whether accurate output can actually be trusted.