Case study
AjustOne
AI-assisted bank reconciliation that escalates to a human instead of guessing.
01
Context
Bank reconciliation in the CEMAC zone is still largely manual: statement labels are inconsistent, and accounting teams spend their closing periods matching lines by hand.
02
Problem
Automate as much matching as possible without ever validating a wrong match. In accounting, a confident mistake costs more than a question.
03
What I did
Product and engineering of the reconciliation engine and its interface (November 2025 – March 2026).
04
Key technical decisions
Deterministic rules first
Exact and rule-based matches are handled by deterministic logic that can be explained line by line.
AI for fuzzy labels only
An AI layer handles what rules cannot: fuzzy matching between inconsistent statement labels.
A calibrated threshold that escalates
Below a calibrated confidence threshold, the match is sent to an operator rather than validated. The system prefers asking to guessing.
Every decision is traceable
Each match records whether it came from a rule, the model or a person, so the whole reconciliation can be audited.
05
Result
90 % of entries are reconciled automatically; the rest is escalated to an operator rather than guessed.
06
Stack
- Nuxt
- TypeScript
- Python
- AI
07
Screens
