Accountants lose 5–10 hours a week chasing clients for receipts and statements. Kontua, our commercial SaaS, automates the whole chase-to-ledger workflow. Here's the architecture: Go + GraphQL core, a Python FastAPI AI pipeline, and a three-database design.
The 5-hour problem
Talk to any small accounting practice and you hear the same story: the work isn't hard, the chasing is. Clients send receipts by email, WhatsApp, sometimes photographs of paper. Staff spend 5–10 hours a week requesting documents and keying data in by hand.
Existing practice-management software solves this — for firms with 50+ staff and enterprise budgets. Kontua is our answer for everyone else, and it's now live with paying customers.
Split by workload, not by fashion
The central architectural decision was to split the system where the workloads differ:
- Go + GraphQL API — the transactional core: clients, deadlines, e-signatures, permissions. Go's concurrency model handles thousands of simultaneous portal sessions cheaply.
- Python FastAPI service — the AI extraction pipeline. Python owns the ML ecosystem, so receipt parsing, invoice line-item extraction, and bank-statement structuring live here, isolated from the core API.
If the AI service is busy processing a shoebox of receipts, the client portal stays fast. They scale independently.
Three databases, three jobs
- PostgreSQL — financial records, where ACID integrity is non-negotiable.
- MongoDB — document metadata, where every client sends different shapes of chaos.
- Redis — job queues and extraction status, where speed matters more than durability.
Everything runs in Docker containers on Alibaba Cloud — and because it's fully containerised, the entire stack can move provider with a day's notice. No lock-in for us means no lock-in for our clients either.
What this proves
Kontua went from idea to paying customers entirely in-house: product design, AI pipeline, backend, infrastructure. When an SME asks us "can you build our platform?", this is the receipt. Let's talk about yours.


