Expense report automation
Receipt photo → ready accounting entry
The task
Large companies share the same pain with expense reports: an employee comes back from a business trip with a stack of receipts and types in the amount, date, seller's tax ID and cost category for each one by hand. An hour and a half to two hours per trip. Then the accountant checks it all again.
Wrong cost categories, typos in tax IDs and lost receipts delay the month-end close. Excel templates don't help: the receipt photos still get copied out by hand.
The client wanted the solution deployed inside its own network: receipt and travel data must not go to someone else's SaaS.
The solution
A platform where the employee photographs the receipt, and the system extracts the fields with OCR + LLM and assigns them to cost categories. Backend in Python, frontend in TypeScript and React. Docker Compose for deployment at the client's site.
The accountant sees a draft report with questionable fields highlighted. One line can be corrected without rebuilding the whole document. Export to a format compatible with the client's accounting system.
Cost categories come from the company's own reference list. The model doesn't invent a category code: it only picks from the list based on the item description on the receipt.
How it works
OCR on the first pass, an LLM on the second for normalisation: the date in ISO format, the amount with kopecks, the seller's tax ID with checksum validation. If confidence is low, the field is flagged for manual review.
Receipt files sit in object storage inside the client's network. The logs contain only the operation ID, with no amounts or names.
Docker Compose: api, a worker for the recognition queue, postgres, minio. Deployed on the client's server in one day with their DevOps team.
Industry context
The need to automate expense reports comes up at companies with 200+ employees and frequent business travel: construction, manufacturing, distribution. 1C can store reports but can't read receipts from photos.
On-premise matters for FinTech and government work: 152-FZ, the ban on foreign clouds, the requirement to audit access.
Security
Deployment inside the client's network, secrets in env, employee / accountant / admin roles. Personal data on receipts doesn't end up in the application logs.
The result
An employee spends minutes instead of hours on an expense report. The accounting team gets structured data instead of photos in a messenger.
How the work went
A pilot in one department with ten employees: we compared manual and automatic reports on the same receipts. Discrepancies in tax IDs and amounts went into improving the prompt and post-processing.
The list of cost categories was loaded from the client's accounting system. The model chooses only from this list; new categories are added by the accountant, not by the neural network.
We put together the Docker Compose setup so the client's DevOps could bring up the stack in one evening: instructions, an env example, a healthcheck on the API.
Roles: the employee uploads receipts, the accountant approves, the admin manages reference lists. The separation is required for internal audit.
For companies with hundreds of business trips a month, the hours saved on reports quickly pay back the development, especially on-premise.
FAQ
Does data go to the developer's cloud? No. Docker Compose is deployed inside the client's network, and receipts and reports stay on its servers.
What if OCR gets it wrong? The field is highlighted, and the employee or accountant corrects it by hand. The system doesn't block the report because of one receipt.
Integration with 1C? Export of structured data. How deep the integration goes depends on the configuration of the client's accounting system.
Which receipts are recognised? Photos of paper and electronic receipts, cash register receipts. Photo quality affects recognition confidence.
Who it's for
Companies with 200+ employees and regular business travel: construction, manufacturing, distribution, consulting. Especially if the accounting team spends days closing expense reports at the end of the month.
On-premise is critical for government work and FinTech: receipts contain personal data and trade secrets. SaaS with servers abroad is often ruled out at the information security review.
Integration with the internal list of cost categories makes accounting entries consistent. Fewer corrections after the chief accountant's review.
Compared with the alternatives
Banking apps can handle receipts but don't know the cost categories of your accounting policy. Here the reference list is your own.
Foreign SaaS for expense management often fails the information security review. Docker on-premise removes the question.
Typing 30 receipts per trip into Excel by hand is a source of errors at the end of the quarter. Automation removes the copying of numbers.
Summary
Employees stopped copying receipts into Excel, and the accounting team gets structured reports inside the company's network. The Docker deployment passed the information security review without external SaaS.
If expense reports hold up your month-end close, we'll go through your process and your list of cost categories. A pilot in one department will show the automatic recognition rate before scaling up.
Working with the studio
We start FinTech projects by agreeing things with information security: where receipts are stored, who sees the logs, how roles work. Docker Compose and deployment instructions are part of the delivery, not an option.
A pilot in one department gives you the OCR accuracy figure on your own receipts. Only after that do we roll it out across the company and connect export to the accounting system.
How to start
Message @vasilk1t on Telegram or write to team@legendastudio.ru. Briefly describe how many expense reports you have per month and your information security requirements for storing receipts. In a 30-minute call we'll tell you whether a one-department pilot makes sense and which receipt formats are recognised best.
For new clients we build a working prototype in 5–9 days with no prepayment: photo upload, field recognition and a draft report on your examples.
The studio is in Saint Petersburg; we support on-premise rollouts remotely or at the client's site by arrangement.
Have a similar task? We'll go through it on a half-hour call. For new clients we build a working prototype in 5–9 days, with no prepayment.