Drone-assisted livestock review for agricultural lending.
Turn aerial photos into cattle observations, reviewable evidence, and a clearer picture of what needs a closer look.
See a review exampleAn independent prototype by Lucas Zheng · 2026
01 / WHY HERDPROOF
I built HerdProof to explore whether drone imagery could give agricultural lenders more context during remote livestock reviews. The question is simple: what was observed, where is the evidence incomplete, and what should a reviewer check next?
The industrialization opportunity is a repeatable inspection workflow: capture, review, and export a record that can be checked again. Its value for lenders still needs to be measured in a real pilot.
02 / THE APPROACH
Keep the original evidence close.
Keep uncertainty visible.
Bring in survey photos, their capture metadata, and the field boundary.
Photos + field contextA cattle detector proposes observations across full-resolution image tiles.
Traceable observationsAlign overlapping photos and merge repeated sightings, including clipped views.
Estimated distinct sightingsInspect uncertain observations, check photo consistency, and export an assessment.
A record + a next actionWhere the walkable scene fits: it adds field context, using measured elevation with approximate photo alignment and cow appearance. It is a visualization, not an independently verified digital twin. Walk the surveyed field
03 / A CONCRETE REVIEW
Imagine a loan officer reviewing a livestock survey. HerdProof's role is to organize the observations and bring unresolved evidence to the surface.
The lending scenario is illustrative. The figures alongside it come from an actual two-photo run of the prototype, completed September 6, 2026.
Download this example recordSURVEY REVIEW RECORD
18 is an estimate of distinct observed animals. It is not a verified herd inventory, ownership check, or loan recommendation.
04 / THE ENGINEERING
Two different tests answer
two different questions.
CATTLE DETECTION
Corrected YOLOv8n v2, evaluated on 77 images at a 0.70 confidence cutoff.
Provisional agreement with 1,041 source annotations. Some labels omit visible cows, and this benchmark had already been inspected. These are exploratory diagnostic results; independent validation is still needed.
OVERLAPPING PHOTOS
The fix compares observations only inside their shared visible area, so a clipped cow can match a fuller view. Unresolved edges remain flagged for recapture.
These tests use supplied annotation boxes. They assess overlap handling, separately from detector accuracy and whole-herd validation.
The prototype does not establish animal ownership, authenticated capture, complete property coverage, loan eligibility, or lender acceptance. No human time-saving claim has been established.
05 / SEE THE MODEL
Saved output from the shipped detector.
Then the same detector, in your browser.
Everything above this point is a number I am asking you to take on trust. Here the detector's actual output is on the photograph, at whatever confidence you choose.
Loading saved detections…
The same 12 MB checkpoint, downloaded into this page and executed on your hardware. Nothing is uploaded, and there is no backend to fall back on — if it does badly, you will see it do badly.
Your photo is read in this tab and never uploaded. There is no server behind this — the model runs on your machine.
Expect it to do badly on most photos. This model was fine-tuned on nadir drone imagery of grazing cattle at survey altitude. A ground-level photo, a different animal, or a different camera angle is outside everything it was trained on, and the result you get will say more about that gap than about the model's measured accuracy.
The same detections, placed on the measured terrain: walk the surveyed field — view only; corrections are made in the local tool.
06 / FROM PROTOTYPE TO PRACTICE
A proposed pilot, with success
defined before the first flight.
An agricultural loan officer or collateral reviewer, working with a ranch and its existing inspection process.
Capture agreed survey areas, review the observations, and export an assessment into the lender's current workflow. A person remains responsible for the decision.
Compare independently checked counts, correction time, missed cattle, recapture frequency, and total review time. Track capture and processing costs.
Independently checked reference data · Secure access to farm imagery · Agreed capture protocol · Measured operating costs