Skip to content
Sycamore ConsultingSycamore
contact@sycamoreconsulting.be
Systems & DataOngoing2026

An indicator ledger where field documents become donor-defensible numbers, with AI doing the reading and humans doing the deciding.

ClientVirunga Foundation
2
Approvals before a figure is reportable
Every change
Audit trail
FR / EN
Interface languages
Pilot, 2026
Status

The Challenge

The park tracks hundreds of indicators across conservation, energy, agriculture and community programmes. The data existed. Reading it did not. Field teams produced reports in every format a field team produces: Word narratives, Excel exports, scanned forms, system dumps. A small M&E team then re-keyed them by hand, and by the time a donor report was due nobody could say with certainty where a given number had come from. The bottleneck was never collection. It was the handful of people who had to turn thousands of documents into figures they could defend in front of a donor.

What I did

I built a centralised indicator ledger with a reporting engine on top. The four-level programme hierarchy is the chart of accounts, editable by administrators rather than hard-coded. Documents land in the shared drive the organisation already uses; staff ask an AI assistant to file them, and it reaches the ledger through an MCP server I wrote, which serves the extraction instructions and validates the result hard. An unknown indicator goes to a mapping queue and is never created automatically. A missing period is rejected outright. Every proposed value then passes two human gates: the uploader confirms the extraction, the indicator owner approves the number. Sensitive material such as security incidents, grievances and ranger casualties sits in a restricted tier visible only to a named allowlist, with aggregates public and detail withheld. Every change is logged and attributable to a person. Reports are drafted from structured templates and saved back as versioned drafts. The interface is French by default with English available, and it runs on plain Postgres with a docker-compose file from the first commit, so the organisation can host it itself.

Outcome

The platform is in pilot on the conservation domain for one full quarter, ending in one real donor report end to end. The change it makes is traceability: any figure in a report leads back to the document it came from and the two people who signed it off. The design point is deliberate and it is what makes the system usable on data donors audit. The AI proposes, it never decides.

Monitoring & EvaluationAI AgentsSoftwareDRCDonor Reporting

Facing a similar challenge?

Write to me.

Get in touch