When small businesses think about accounting, they rarely imagine distributed systems, ML pipelines, and polymorphic ledgers. For Peter Potapov, that’s exactly where the work starts.
Over the last decade, Potapov has gone from founding his own analytics startup to leading engineering and machine learning at Osome — the Singapore-headquartered fintech behind one of the first “Accounting Factory” models — and now to driving AI-first finance products at Kick in the US. His focus has been consistent: turn messy, incomplete financial data into reliable, auditable books at scale.
Joining Osome: Rebuilding Accounting From the Ledger Up
When Potapov joined Osome in Singapore as Head of Engineering & ML, the company had already raised venture funding and was serving thousands of SMEs across Singapore, Hong Kong, and the UK. Osome handled incorporation, accounting, payroll, and tax reporting via a digital-first platform — but its leadership wanted something more ambitious than another layer on top of Xero or QuickBooks.
They were betting on a hybrid “Accounting Factory”: a production system where machine learning did the heavy lifting on document processing, categorization, and reconciliation, and human accountants focused on judgment calls and edge cases. Potapov was hired to architect and deliver that system.
A Polymorphic General Ledger for Three Jurisdictions
The first problem was foundational: the ledger itself.
Instead of building a tangle of country-specific schemas, Potapov led the design of a polymorphic document/transaction model capable of expressing all major business operations in a single coherent structure.
On top of that core model, his team added the machinery needed for production accounting at scale: support for prepayments and overpayments that could be unwound deterministically, payroll workflows that respected local rules while feeding a unified ledger, logic for deferrals and accruals so revenue and expenses landed in the right periods automatically, plus automatic warehouse and COGS calculation, real-time reporting, and robust lock management.
The result was a general ledger tailored to the realities of SMEs in three regions (UK, SG, HK), but engineered as a single extensible system rather than three separate codebases.
Turning Noisy Inputs Into Reliable Books
Once the ledger was in place, the real battle began: data.
In his article “How We Built an Efficient ML Model With Dirty Data and Insufficient Information,” Potapov describes the central challenge bluntly: you don’t get clean datasets in bookkeeping; you get exactly the opposite, and you still have to ship models that work.
The Human-in-the-Loop Accounting Factory
To deal with this, he designed and led an end-to-end automated accounting pipeline:
document / transaction → enrichment → categorization → reconciliation → human QA → learning loop
ML models pre-labeled transactions and documents, assigning likely categories, counterparties, and tax treatment before an accountant ever touched them. High-confidence cases flowed through a fast-track review loop; ambiguous or high-risk items were routed into more detailed workflows with richer audit trails.
A key innovation was counterparty normalization: collapsing messy vendor strings into stable entities through deterministic rules and ML-based similarity, dramatically improving downstream matching. Bank transactions, invoices, and other documents were reconciled using a layered approach — strict deterministic rules first, machine-learning similarity and heuristics second, human review only where necessary. Every correction from accountants fed back as labeled data, closing the learning loop.
For accountants, the impact was concrete: in many flows, the time spent on manual categorization and reconciliation dropped by as much as half, while accuracy and consistency improved because the system treated similar cases the same way every time.
Crucially, the point of this design was not to slow automation down, but to let models do the mechanical work at scale while humans injected client-specific context and domain judgment that would never fully fit into a training dataset — and to capture that judgment back into the system as structured feedback.
Reliability at Scale: Three Markets, Thousands of Entities
Accounting is not a “move fast and break things” domain. Late or incorrect books have real regulatory and tax consequences.
As Osome expanded, Potapov’s systems had to support three active regions (Singapore, Hong Kong, UK) with different tax codes, filing cycles, and reporting standards; thousands of client entities; and tight SLAs for month-end and year-end closes.
To keep the Accounting Factory reliable under that load, he and his team implemented multi-region production designs that balanced cost, latency, and model placement; staged enrichment and selective model routing so heavier ML workloads only ran where they actually moved the needle; and targeted human verification focused on high-value, high-risk items. The system operated at 99%+ availability while producing records suitable for regulated reporting, not just internal dashboards.
Beyond Osome: Continuing the Work at Kick
After his tenure at Osome, Potapov moved to the US and took on a leadership role at Kick, an emerging AI-native finance platform. There, he continues to work at the intersection of accounting, AI, and product design — this time with a focus on US businesses and a new set of banking, payroll, and tax integrations.
The problems rhyme: fragmented financial data, small teams drowning in back-office work, and founders who never started their company to become part-time accountants. The solutions build on the same principles he proved out at Osome: ledger-first system design rather than bolt-on “AI features”; human-in-the-loop workflows that make professionals faster instead of trying to replace them; and cost- and latency-aware model architectures, so AI features can run continuously in production instead of only in demos.
What Ties the Career Together
Looking across Potapov’s career, a few patterns stand out. He treats accounting systems as mission-critical infrastructure, not just SaaS dashboards. He builds ML/AI architectures that assume dirty data and partial information, instead of waiting for perfect inputs that will never arrive. He insists that ML/AI pipelines stay wired into real user workflows, context, and domain expertise that sit outside any model’s training distribution — using human-in-the-loop routing, audit, and feedback to inject that knowledge — rather than throwing ungrounded “black box” models into production. And he has operationalized these ideas at real scale, across multiple regions and thousands of companies, with uptime and auditability that hold up under regulatory scrutiny.
For regulators and investors, that combination — deep technical ownership, verifiable production impact, and influence on how others build similar systems — is what separates routine engineering from work that genuinely moves a field forward.
