The in-house software stack

This is the part that decides whether the business works. Hardware is a commodity with a known price. Processing and analytics are normally rented per hectare per year, which means your best customers are also your biggest cost line, forever.

None of it has to be rented.

The pipeline, end to end

flight plan ──▶ aircraft ──▶ raw frames + flight log
                                   │
                          ┌────────┴────────┐
                          ▼                 ▼
                    geotagging        radiometric
                    (RTK + trigger)   calibration
                          └────────┬────────┘
                                   ▼
                    ┌──────────────┴──────────────┐
                    ▼                             ▼
             photogrammetry                  per-frame
             (ortho, DSM, DTM)               inference
                    │                             │
                    ▼                             ▼
              raster layers ──────────▶ detections in field space
                    │                             │
                    └──────────────┬──────────────┘
                                   ▼
                          clustering + zoning
                                   ▼
                    ┌──────────────┴──────────────┐
                    ▼                             ▼
             prescription export            flight record
             (ISOXML, shapefile)            (this repository)
                                                  ▼
                                        site, API, dashboard

The five pages

What to build and what to adopt

The pattern: adopt the mathematics, build the judgement. Nobody needs to write a bundle adjuster. Everybody needs to decide what area threshold switches a nozzle on, and that decision is the product.

The rules that keep the stack honest

These come from the same place the rest of this codebase comes from, and they are worth stating as engineering rules rather than as taste.

Every figure is derived, never stored. Store the records; compute the totals on read. A stored total drifts away from its evidence and the audit trail becomes decorative.

Every output carries its provenance. A detection knows its flight. A prescription knows its detections. A report knows its prescriptions. The field report shape in this repository is the example: treatedHectares comes from a TreatmentMap row, not from copy.

Determinism where it is possible. The demonstration field is generated from a fixed seed with no Date and no Math.random, so the database, the API and the rendered map always agree. Apply the same discipline to processing: the same frames and the same model version must produce the same detections, which means pinning model versions into the flight record.

Sample data announces itself. /api/field-report carries sample: true in the payload rather than in a footnote. Anything modelled, estimated or interpolated should say so in the data, not in the documentation.

Fail one panel, not the page. Every read in this repository is wrapped so a missing table returns a fallback. The processing pipeline deserves the same: one failed band, one unmatched frame, one missing log should degrade an output, not lose a flight.

Total software cost

The trade is honest: you exchange a permanent per-hectare cost for a one-off engineering cost and an ongoing maintenance obligation. It is the right trade above a few thousand hectares a year and the wrong one below a few hundred.