ClearLedger — AI-Powered Analytics Dashboard

Industry: FinTech / Accounting SaaS Service: Analytics platform + LLM integration Timeline: 8 weeks Stack: Next.js, TypeScript, Anthropic, Supabase, Recharts, Vercel

---

The problem

ClearLedger had a strong backend — years of structured financial data in a well-maintained Postgres warehouse. The problem was getting that data in front of the people who needed it.

Their CFO customers were technically capable, but they weren't SQL writers. Every insight required a ticket to the data team, a day or two of wait time, and a spreadsheet export that was stale by the time it arrived. ClearLedger's NPS sat at 31, and exit interviews consistently pointed to the same thing: "I can't get to my own data without asking for help."

What we built

Analytics dashboard layer: A Next.js application layered directly over the client's existing Postgres warehouse. Server-side queries, chart components built with Recharts, and a role-based access system so each CFO customer saw only their data.

"Ask your data" feature: An LLM-powered query interface backed by Anthropic's Claude. Users type questions in plain English — "show me operating expenses by category for Q1 vs Q2" — and receive rendered charts with source data. We built a structured query generation system that converts natural language to parameterized SQL, keeping the model away from raw query execution.

Saved queries and scheduled reports: Users can bookmark any chart or table, set a refresh schedule, and receive the output via email. Reports that previously required manual compilation now run on autopilot.

The technical decisions

Anthropic over other providers: The structured output quality for SQL generation was meaningfully better in testing. ClearLedger's data schemas are complex (multi-currency, multi-entity consolidation), and Claude handled ambiguous column semantics more reliably than alternatives.

Parameterized SQL over raw LLM queries: We never let the model write or execute raw SQL. Instead, it generates a structured intermediate representation — a query intent object — which a deterministic layer converts to parameterized queries. This eliminates injection risk and makes the system auditable.

Recharts over a heavy BI library: The client's CFO customers wanted speed, not a Tableau replacement. Recharts gave us lightweight, composable charts that rendered fast on any connection. The goal was a 10-second insight, not a 10-minute dashboard build session.

Supabase for user management: Rather than building a parallel auth system, we used Supabase's row-level security to enforce customer data isolation at the database layer. Each CFO org gets a policy. No application-level filtering required.

Results

• NPS: 31 → 58 within three months of launch
• "Ask your data" adoption: 74% of active accounts use it weekly
• Excel export usage: Dropped 83% — customers stay in-product instead of exporting to analyze elsewhere
• Support tickets: Data access requests to the data team down 67%

Handoff

Full codebase delivered in the client's GitHub organization. Supabase project transferred to their account. Includes schema documentation, LLM prompt library with versioning, and a runbook for adding new chart types. 30-day bug warranty included.