A proof-of-work built for Minerva — the YC-backed startup automating accounting for small businesses. The core problem: accountants spend hours manually mapping raw bank transactions to General Ledger codes. This app does that with AI — and puts the accountant in control of every decision the model makes.
RI
Minerva was founded by Peter Zhu and Om Agarwal, backed by Y Combinator. They're building AI for the painful parts of finance — automating the workflows that traditional accounting firms run manually today.
The mission: bring AI-powered automation into accounting firms they partner with and acquire, reducing the time accountants spend on low-judgment work so they can focus on what actually requires expertise.
View on YCUpload
Accountant uploads any bank export — CSV, XLSX, or XLS.
Schema Detection
Heuristics map column headers first; falls back to Gemini 2.5 Flash for non-standard formats. Zero API cost on most uploads.
Normalize
Dates, amounts, and missing fields are repaired before categorization.
RAG Retrieval
Each transaction is embedded and matched against the GL knowledge base via pgvector cosine similarity (top-3 chunks per transaction).
Categorize
Batches of 20 transactions + retrieved context are sent to GPT-4o in a single prompt. Returns category, confidence, and reasoning for each.
Review
Accountant approves, overrides, or skips. Every action is audit-logged immutably.
Transaction
Core record with date, vendor, amount, AI category, confidence, reasoning, and status (PENDING / APPROVED / OVERRIDDEN).
Upload
Tracks each imported file — filename, transaction count, and processing status. Transactions FK to their upload.
AuditLog
Immutable log of every approve and override action, with old and new category values.
KnowledgeChunk
GL accounting knowledge stored as 1536-dim embeddings (pgvector). Queried via cosine similarity at categorization time.
Client
Represents an accounting client. Transactions are scoped per client for multi-firm support.