Case study · AI & Data Pipeline

Accountant Bot: paper invoice parsing and intelligent data routing.

Replaced manual invoice logging with an intelligent AI pipeline. The manager simply drops a photo of a crumpled invoice in Telegram — Gemini parses the table data, normalizes measurement units (e.g., milliliters to liters), and automatically routes the records and scans into Google Drive and Google Sheets.

ProjectDocument Workflow Automation RoleAI Engineer · Backend Development Key FeatureVision API (OCR) + Supabase + Google API
99% paper invoices
digitized automatically
0 manual inputs to
spreadsheets
Auto automatic unit
normalization
01 / Context

The headache of paper invoices

In Montenegro, less than 1% of vendors use electronic document exchanges (EDI). Nearly all stock delivery to restaurants and stores arrives with printed paper invoices.

To log a product delivery into the system, an accountant or manager must sit at a PC and copy-paste each item name, quantity, and price into endless Google Sheets. It is a grueling, repetitive chore that eats up hours and inevitably leads to typos, inventory mismatches, and discrepancies.

No Bot / With Accountant Bot

— Before
  • Hours of manual data entry from physical invoices
  • High human-error risk (price typos)
  • Document photos lost in messy group chats
  • Unit mismatches: vendor bills 750ml, but database requires liters
+ After
  • Snap an invoice photo and drop it to a Telegram bot
  • AI parses table text and returns a clean PDF draft for review
  • Bot automatically converts 750ml to 0.75l for clean inventory metrics
  • Archiving: scans upload straight to target Google Drive folders
02 / Under the hood

Parsing, validation, and routing

We built a Serverless architecture powered by Supabase and Gemini Vision API. The entire document ingestion pipeline finishes in seconds and is guarded against LLM hallucinations with human-in-the-loop validation.

  • Step 1: Parsing Vision API. The user snaps and sends an invoice photo to the bot. Gemini processes the image, reads the table structure, and extracts raw data.
  • Step 2: Validation Review. The bot compiles the parsed rows into a clean PDF preview and sends it back to the user. The manager compares the PDF with the physical paper and taps "Confirm" if it matches.
  • Step 3: Routing Google Drive. The bot automatically creates a dedicated archive folder on Google Drive and uploads both the original photo and the generated PDF.
  • Step 4: Normalization Data Pipeline. The records are pushed into shared Google Sheets. Simultaneously, the bot normalizes metrics: if a vendor lists 750ml, the bot converts it to 0.75 liters so warehouse tracking balances correctly. Sheets are compiled in both Russian and Montenegrin.
We turned a crumpled piece of paper into a structured, normalized database record, saving the client dozens of hours of manual labor every month.

Need similar workflow automation? Tell me about your task — I’ll draft what is feasible.

Message on Telegram
Back to cases