Hi all,
I'm Raman Lamba. I've built a small open-source tool on top of Fineract and
would value your feedback on how it uses the API.
What it does
It matches incoming bank payments (CSV, camt.053 or MT940) to loans in
Fineract and posts them as repayments. For co-lending setups it also
records the lender/partner split. Fixed rules decide every payment.
Anything ambiguous goes to a review queue for a person. An optional AI
model can help read unfamiliar statement layouts, but it never picks a loan
or posts anything.
How it uses Fineract
- It reads loans and repayment schedules, and its only write is POST
/loans/{id}/transactions?command=repayment.
- It sends the bank's UTR as the transaction externalId, so Fineract's
uniqueness check blocks double posting. A crashed run recovers through GET
/loans/{id}/transactions/external-id/{id}.
- It runs against the openMF mcp-mifosx sandbox. It never creates or
changes clients, loans or products.
Questions for you
1. Is using externalId as the idempotency key for repayments the right
approach? Is there anything in Fineract's existing batch or reconciliation
features I should use instead?
2. Is a least-privilege user with only READ_LOAN and REPAYMENT_LOAN (plus a
few read permissions) enough? I haven't tested that yet.
3. Fineract has no proportional co-lending split, so I keep the split in an
external ledger. Is there a recommended way to model it?
Limits
All data and all measured results are synthetic (0 wrong postings in 1,348
automatic posts, but on data the same project generated). No real bank
export has been through it, and no live AI call has been made. It is not
production-ready.
Link:
https://github.com/Raman-Lamba/fineract-loan-reconciliation-AI-supported-agent
(MIT licensed; python -m ui.serve --demo needs only Python 3.11+)
If you run Fineract in production and can try it against a test instance,
or share an anonymised bank export format, that would help most. Thanks for
your time.
Raman Lamba
https://ramanlambaportfolio.vercel.app