What is the problem your project solves?

Personal finance software is built on a false premise: that your bank feed is the truth and your documents are clutter. The consequences are everywhere. A single credit-card purchase produces a merchant invoice AND a bank confirmation — and every tool counts that rupee twice. Apps silo themselves: one does bank accounts, another mutual funds, a third stocks, so none can tell you what you're actually worth. And the price of admission is standing access to your email and bank accounts.

Quick2AVault inverts the premise: personal finance is a document-reconciliation problem, not an aggregation problem. You feed it documents — a watched folder, an email dropbox, an API — and it builds a deduplicated, reconciled, evidence-backed ledger of your economic life. On your machine. No standing access to anything.

How you are solving it?

A headless TypeScript daemon runs a three-phase pipeline: SHA-256 intake with dedupe → deterministic markdown conversion → Claude extraction into a strict JSON schema. The canonical schema models any economic event — who paid whom, how much, why, when, how, with what evidence — using four entity kinds (person/organisation/account/instrument), double-entry transaction legs, and documents-as-evidence, with invariants enforced as database constraints (integer minor units; transfers structurally excluded from spending). A scored reconciliation engine links invoices to bank alerts: one rupee, counted once.

Three clients share one versioned Core API: a web UI, a native macOS menubar app, and an MCP server — so an AI agent can query your finances with full provenance. Jurisdiction packs (India + Japan shipped) make every country data, not code.

Proven on 69 real documents → 67 transactions, 0 failures — including catching a ₹1,68,641 salary misbooked as a transfer, and separating ₹2,05,943 of investments from honest spending of ₹40,283.

Disclosure: the repo root contains a pre-existing macOS prototype (Glaze runtime) and prior architecture plans. Everything under daemon/ and desktop/ — the daemon, schema, extraction, reconciliation, MCP server, both UIs, jurisdiction packs, Learning Mode — was built in the hackathon window on branch hack/push-to-prod.

How Did You Use Claude?

Claude is the judgment layer, not a feature. Its only product is structured truth: strict tool-schema JSON extracted from canonical markdown (determinism via forced tool_choice — Claude 5 rejects the temperature parameter). Claude drives entity resolution (knowing a wallet-load is a transfer, not spending; that Swiggy the merchant, Swiggy the equity, and the Swiggy wallet are three different entities) and the reconciliation semantics that collapse duplicate evidence into one transaction.

The day's hardest bug was an AI-semantics bug — the extractor invented accounts and booked a salary as a transfer — fixed by bounding model authority with code invariants: a prompt is not an invariant. User-confirmed values are never overwritten by AI. And through the MCP adapter, Claude becomes an interface to the vault: ask "what did I spend this FY?" and get an answer with evidence links.

What is the deployed URL for this project?

Haha trick question, right? There is no public URL. Quick2AVault is local-first by design.