SlipMind

The messy workflow

Sole traders lose financial proof constantly. A receipt is a crumpled photo; an invoice is buried in a forwarded email; a payment is a half-remembered line in a bank feed. By the time accounting software or an accountant sees it, context is gone and evidence is incomplete — which means time lost, claims missed and stress at year-end.

The gap isn't accounting. It's the layer before accounting: capturing and structuring evidence early enough to make it useful.

Inputs

• Receipts and invoices (photos, PDFs)
• Forwarded emails
• Manual quick entries
• CSV exports from banks and tools

Output

• Reviewed, structured records
• Missing-evidence flags
• Record hygiene score
• Accountant-ready CSV exports

System flow

The AI layer — and its limits

AI does extraction and classification: reading unstructured evidence, proposing fields, suggesting categories and attaching a confidence level. It does not make final accounting judgements and never auto-submits. Every record passes through a human review step before it can be saved or exported.

Guardrails

• Human review required before save or export
• Confidence scores surfaced on every extracted field
• Missing-field and missing-evidence warnings
• Source-linked output — each record traces back to its evidence
• No automatic final submission
• CSV-injection protection on export

Status

Working prototype with a public demo repository on GitHub → showing the core pattern end to end — messy evidence in, quality-scored record out — with example data only. Private walkthrough available during interview.

What I learned

The hardest part of an evidence product isn't extraction accuracy — it's making uncertainty visible without making the user do accounting. Confidence scores and a hygiene score did more for trust than any model improvement, because they told the user exactly where to look. The system earns trust by admitting what it doesn't know.