Skip to content

Private crop-model evaluations

Private evaluations for crop AI.

We sell private evaluations of crop growth, plant health and harvest timing for agricultural AI teams. We test a model against reviewed observations and report where it succeeds, where it misses and where the evidence is too thin. Our ten years of raw field data stay in-house.Private tests of crop growth, plant health and harvest timing. Your team receives findings. Our ten years of raw data stay in-house.

First commercial milestone: a paid evaluation with an independent team. No outside customer yet.

Conceptual visualization · no live inference

Ten years of curated field data.

MicroHabitat has collected field photographs, growing records and horticultural knowledge since 2016. We use this private material to design tests. Each evaluation draws on a dated, reviewed subset that fits the crop and the question.

Explore the model evidence
A MicroHabitat grower tending crops in rooftop planters
Field care · observations in context
Tomato plants showing fruit and foliage
Crop detail · fruit, foliage and structure
Hands working with plants in a MicroHabitat garden
Planting · the work around each crop
A MicroHabitat rooftop growing environment
Growing environment · a view of the field
Existing source
16,244field photographs196 farms · 3 Nov 2026

A growing field collection

The MyUrbanFarm collection complements ten years of curated field archives.

Information to connect

Connect the visits

Link the observed condition, a dated care action and what changed next. Planter identities and sequence links still need verification.

  1. PhotoObserved condition
  2. Care actionDated intervention
  3. Later visitChange over time
Research hypothesis

Can history improve predictions?

Compare both approaches to find out whether verified history improves predictions. The gain is not yet established.

BaselineImage only
Research variantImage + verified history
See the research evidence

Archive photographs illustrate the field context. Each model evaluation uses a dated subset. Matched planter sequences are still being verified.

Private evaluation service

What a research team receives

  • Test protocol
  • Aggregate scores
  • Error analysis
Discuss a private evaluation

A decision the team has to make

A lab might need to choose between two models before deploying one in a greenhouse. We agree on a crop, decision and baseline, then test both models on the same held-out observations inside our private evaluation environment.

Results the team can act on

The report gives the protocol, aggregate scores, error patterns, uncertainty and coverage limits. A repeat evaluation can test a revised model against the same criteria. We quote each engagement after confirming that the required observations exist.

Tomato · Rooftop D

The record we are building

Synthetic example of the target format. Real planter-level action and outcome sequences are still being assembled.

  1. Week 1

    aphids

    Action · treated
    First record
  2. Week 2

    aphids, fewer

    Action · treated
    unchanged
  3. Week 3

    no problem found

    Action · pruned
    recovered
  4. Week 4

    no problem found

    Action · watered
    unchanged
  5. Week 5

    leaf scorch

    Action · watered
    worse
  6. Week 6

    no problem found

    Action · harvested
    recovered
  7. Week 7

    no problem found

    Action · harvested
    unchanged
  8. Week 8

    no problem found

    Action · no action
    unchanged

Fictional example · one planter over time

Capture software makes future records linkable

  1. Observe

    Photograph a planter and record the observed condition.

  2. Record the action

    Attach the care action to that planter, including when no action was taken.

  3. Revisit

    Compare a later observation. Record the change without assuming that the action caused it.

What exists, and what is next

Existing source operation
Farm photos, issue reports and farm-level follow-ups are recorded in MyUrbanFarm.
Working review demo
Try planter selection, visit logging and follow-up review with synthetic records. The demo saves locally in your browser.
Next data milestone
Verify planter identity across visits and have experts label observations and visible changes.
Try the capture workflow

From crop models to farm robots

We offer private evaluations to robotics teams. Our next products would let manufacturers license our own crop intelligence and growers buy Farm Bot, the machine designed to put it to work.

Explore our vision
Farm Bot concept with its receiving and solar stations

Our model tests the data thesis

Our current model trails the hosted baseline.

46% versus 80% correct problem names on 50 human-judged problem photos from 21 held-out farms, measured 27 October 2026. This is a small, selected sample of photos the hosted model flagged; missed problems are not measured.

This classifier establishes a starting point. Next, we will test whether verified care and follow-up records improve predictions beyond a single image.

Read the benchmark
Our model46%
Hosted frontier model80%
Problem named correctly · 50 photos, 21 farms95% interval of the gap: 15–56 points · Measured 27 Oct 2026

The source of our field data

develops crop models and tests how reliably they perform on field observations.

About the company

Field roots

Experience since 2016

The team brings engineering and field experience from MicroHabitat’s network, established in 2016.

Company relationship

A related-party starting point

MicroHabitat Inc is the intended first related-party customer and minority shareholder.

Next milestone

An outside design partner

The next commercial milestone is to work with an independent team on a defined evaluation.

Website assistant

What would you like to know?

Ask about our crop AI, Farm Bot or field evaluations.

Enter to send · Shift + Enter for a new line

Anthropic receives your question, relevant earlier messages from this conversation and relevant source passages. Privacy