Seven AI models across 300K hectares — at 96%+ precision.
300K+ hectares · 7 AI models · 96%+ precision
Read case →02 · The challenge
Millions of aerial frames a season. 460+ weed species across nine global regions. Previous vendors could move the volume — they couldn’t hold the taxonomy across seasons.
Each frame demands species-level identification with sub-1% misclassification — early enough in the growth cycle to drive prescription decisions.
Annotator churn at previous vendors meant the same taxonomy got re-learned every quarter. Accuracy degraded under load.
03 · How we did it
A taxonomy-first programme with the same team across four growing seasons.
01
Taxonomy
Built the 460-species weed ontology with Taranis’ agronomy team. Boundary rules for ambiguous early-stage growth. Gold set curated in week one.
Schema-first · 460 species ontology
02
Multi-pass review
Three-level QC across every batch — L1 annotator, L2 reviewer, L3 QA specialist. Disagreement-resolution protocol for ambiguous frames.
Three-pass review · disagreement protocol
03
Proactive tooling
Internal dashboards built before the customer asked. Inter-rater agreement, throughput, and seasonal drift flagged automatically.
Proactive QA dashboards · drift detection
04
Capacity scaling
50% annual capacity growth without quality drop-off. Same delivery lead, expanded annotator team — knowledge compounded inside the team.
50% YoY capacity growth · zero quality drop
04 · The outcome
The accuracy held because the team didn’t change. A 70-person expert team with domain depth across 30 crops — same annotators, same QA leads, year over year. Knowledge compounded inside the team.
Taranis scaled annotation capacity by 50% year-on-year without losing precision. The proactive QA dashboards we built — not because Taranis specified them, but because we knew proactive QA was the surest way to drive accuracy and performance on the programme — flagged drift before it reached production.
05 · In their words
“IndiVillage is the only one of our suppliers doing their own internal QA. The outcome is superior to the rest.”
Idan Harary · Director of Ag · Taranis
06 · The numbers underneath
460+
Weed species
in live taxonomy
4.5M+
Images annotated
to date · sub-1% miss
3-level
QC review
on every batch
07 · Other programmes that shipped
Same operating discipline. Different modalities.
08 · Work with us
Run an annotation audit on your data.
Send us 100 frames in any modality — image, video, LiDAR, audio, text. We'll return annotated output, an accuracy benchmark, and a programme recommendation in 48 hours.