Lucid - AI-Powered Expense Intelligence

Inspiration

SMBs on corporate cards generate thousands of transactions monthly but lack the intelligence tools that enterprises take for granted. Finance managers are stuck manually reviewing spreadsheets, chasing policy violations over email, and building expense reports by hand. We saw Brim's real anonymized transaction data and asked: what if a finance team could just talk to their spending data and have AI proactively surface what matters?

What it does

Lucid is an AI-powered expense intelligence platform built on 6 months of real anonymized Brim card transaction data (4,180 transactions, 9 cardholders, 3 departments). It delivers four core capabilities:

Talk to Your Data — Finance managers ask questions in plain English ("What did Fleet Operations spend on fuel vs permits last quarter?") and get back charts, tables, and summaries with full multi-turn context
Policy Compliance Engine — Digitized Brim's real expense policy with contextual AI analysis. A $200 solo dinner flags a violation; the same $200 dinner with 6 attendees passes — the AI explains why
AI Pre-Approval Workflow — Pending transactions surface with employee spend history, department budget status, and an AI-generated approve/deny recommendation with reasoning. One click to decide
Automated Expense Reports — AI-generated reports with category breakdowns, policy compliance checks, and narrative summaries ready for CFO review

Beyond the requirements, we built an AI Autopilot Advisor that autonomously chains 5 tool calls (KPIs, approvals, violations, anomaly detection, budget health) through multi-step reasoning to produce a prioritized action plan — streamed in real-time through a terminal UI. We also built spend forecasting with linear regression, vendor consolidation analysis, and employee peer benchmarking.

How we built it

Next.js 16 with Turbopack for the frontend and API routes
SQLite as the data layer — synchronous queries against 4,180 real Brim transactions
Claude API (Sonnet) for multi-step agentic reasoning with tool use — the agent loop streams results so users see AI thinking in real-time
Groq and DeepSeek (via OpenRouter) as alternative model options
Recharts for data visualization — bar charts, pie charts, area charts, all with dark-themed custom tooltips
Clerk for authentication
Pattern analysis on 500+ merchant names to break down fleet operations into fuel, permits, tires/parts, and maintenance subcategories
Real Brim expense policy digitized into structured rules for the compliance engine

Challenges we ran into

Groq rate limits hit us mid-demo prep. We implemented dual-key rotation with automatic fallback and added DeepSeek as a third model option via OpenRouter's API using native fetch with manual SSE parsing
Fleet category granularity — the raw data had all 3,500+ fleet transactions lumped under "transportation." We analyzed 500+ distinct merchant names and wrote SQL pattern-matching rules to classify them into 4 meaningful subcategories, making the dashboard actually useful for fleet cost management
Multi-step agentic reliability — getting the AI Autopilot to consistently call all 5 tools in sequence and produce a structured action plan required careful prompt engineering and a robust streaming agent loop. Some models (like DeepSeek) couldn't handle the tool chaining, so we route Autopilot exclusively through Claude
Real data, real messiness — anonymized card numbers, inconsistent merchant naming, mixed transaction types. We had to filter credits/point redemptions and build a complete data pipeline from XLSX to JSON to SQLite

Accomplishments that we're proud of

Contextual compliance — the system understands that a $200 team dinner and a $200 solo dinner are fundamentally different policy situations. This is not threshold matching; it's reasoning
The Autopilot terminal — watching the AI autonomously scan KPIs, check violations, detect anomalies, and produce a prioritized action plan in real-time is genuinely impressive. Multi-step agentic workflows, not single-prompt wrappers
Fleet cost discovery — breaking "transportation" into fuel ($731k), permits ($359k), tires ($232k), and maintenance ($47k) was a genuine insight from the data that a finance manager would immediately act on
Three-model architecture — Sonnet for depth, Groq for speed, DeepSeek as a free fallback. Users toggle with one click

What we learned

AI depth beats feature breadth — two features that work beautifully create more impact than six that half-work. We focused on making compliance contextual and the autopilot truly agentic rather than adding more screens
Real data changes everything — working with actual Brim transactions forced us to solve problems we'd never encounter with synthetic data: merchant name disambiguation, category inference, handling anonymized employee IDs
Prompt engineering is architecture — the difference between a single-prompt wrapper and a multi-step agent is not just more API calls, it's designing tool schemas, managing conversation context across turns, and handling partial failures gracefully
Streaming matters for trust — showing the AI's thinking process in real-time (tool calls scrolling through the terminal) builds user confidence in a way that a loading spinner never could

What's next for Lucid

Receipt matching — OCR integration to auto-match uploaded receipts against transactions
Slack/Teams integration — push approval requests and violation alerts directly into team channels
Multi-company support — let Brim offer Lucid as a white-label intelligence layer for all their SMB clients
Predictive policy violations — flag likely policy breaches before the transaction clears, using spending pattern analysis
Custom policy builder — let finance teams define and modify expense rules through natural language instead of configuration forms

Built With

claude-api, next.js, react, typescript, tailwindcss, sqlite, better-sqlite3, recharts, clerk, groq, deepseek, openrouter, motion, turbopack