Receipt management built for Europe's independent consultants.
The Problem
Every independent consultant in Europe loses hours a year to the same three things: receipts scattered across photos, PDFs, and inboxes; VAT rules that differ by country; and exports their accountant can't actually use. ReceiptHub fixes all three in one flow — and runs entirely on the user's own machine.
How It Works
Five steps between a paper receipt and a clean, tax-ready record. The receipt image never leaves the user's device.
Tesseract.js runs OCR directly in the browser. Merchant, amount, date, currency, and country all get pulled out on-device. The receipt image never leaves the user's machine.
On-device OCRA 119-chunk knowledge base (27–32 rules per country, written by tax professionals) is searched with BM25 — the same scoring family behind Elasticsearch — expanded with multi-language synonyms so a French "repas" and a German "Bewirtung" both match the right rule.
BM25 retrievalThe top 6 matching rules go to Llama 3.2 3B, running locally through Ollama. The model has to name the type of expense first, then pick which rule applies, before it commits to a category. That ordering stops a small model from grabbing whichever rule scored highest regardless of fit.
Local LLMA rule-based post-processor checks the output against known failure patterns before the user sees it — a parking ticket wrongly filed as Accommodation, a deductibility value returned as a fraction instead of a percentage. Every correction is logged next to the model's original answer, so a reviewer can always see both and revert if needed.
Sanity checkerVAT rate and statutory basis, audit-risk flags, the exact income-tax citation for that country and category, and a documentation checklist — all computed from the classification, no second AI call needed.
DeterministicPhase 1 · MVP Features
No bloat. The Phase 1 MVP is scoped tightly around the core job-to-be-done: capture a receipt, understand what it is under local tax law, make it exportable.
Extracts vendor, amount, currency, date, and line items from any receipt format — photo, PDF, or email attachment — in seconds.
AI categorisation pipeline built on a versioned RAG system — tax rules change, and so does the model, without breaking existing records.
Separate rule sets for Germany (DE), France (FR), Netherlands (NL), and Spain (ES) — each with country-specific expense categories and deductibility criteria.
Export everything in open formats: CSV, PDF, or DATEV. No lock-in. Your accountant gets clean, structured data; you stay in control.
Market
The EU has one of the world's largest and fastest-growing independent workforce populations. ReceiptHub targets the highest-value segment — consultants and freelancers earning above median — in four markets with strong VAT complexity.
Launch Markets
Each country has distinct VAT rules, rates, and expense classification requirements. Rather than building a generic EU app and hoping it fits everywhere, ReceiptHub ships with country-specific tax logic baked in from day one — making it genuinely useful rather than a glorified spreadsheet.
Pricing
One plan, one price. No tiers, no hidden usage limits, no "export is a premium feature" nonsense. At €99/year, ReceiptHub costs less than a single hour of the accountant time it saves.
Technology
The core insight: off-the-shelf LLMs hallucinate on jurisdiction-specific tax rules. ReceiptHub uses a versioned RAG pipeline so the AI's knowledge base can be updated as rules change — without retraining the model.
A cloud model would be more accurate out of the box, but it costs per receipt, sends every receipt to a third party, and needs a live connection. Running Llama 3.2 3B locally through Ollama costs nothing per call, keeps data on-device, and works offline. The accuracy gap is closed by the RAG retrieval and the sanity-checker — not by a bigger model.
Embeddings would be more semantically sophisticated, but they need a hosted embedding model or a heavy local one. At 119 chunks across 4 countries, BM25 with multi-language synonym expansion does the job in about 3 KB of code — no vector infrastructure required.
Tax law changes every year. Fine-tuning would mean a retraining cycle per country per year and an ML pipeline to maintain. With RAG, a tax attorney edits a markdown file, one script rebuilds the corpus, and the system is current the same day. No ML pipeline.
Extracts structured fields (vendor, amount, date, currency, line items) from any input format. Confidence scores surfaced to user for review and correction.
Every user correction feeds back into personalised categorisation patterns. The more you use it, the more it learns your specific expense profile and vendor habits.