Supporting evidence
Crop-health AI
The complete artifacts and detailed records behind the account. Current public references, original work and proposed designs retain their labels.
Detailed material
The problem
A diagnosis product had to work where the farmer needed help, including on constrained phones with uneven connectivity. A crop label alone would not resolve the practical question of what to do next.
Capture evidence without waiting for a network
The delivered flow let a farmer photograph a leaf offline and receive the diagnosis when the phone reconnected. That separated evidence collection from connection availability, rather than require both to happen at the same moment.
The product question extended beyond whether a model could identify a condition. A usable journey needed an obtainable photograph, a path back to the answer and a clear boundary around what that answer established.
Work across data, models and agronomy
The project included image curation, labelling and an agronomist review loop. Specialists helped establish the meaning of labels while engineering and modelling colleagues delivered the technical capability.
My product remit connected that work with the farmer’s help-seeking flow: what evidence could be collected, when an answer could return and how the result could be understood. It does not imply that I personally trained the models or created every annotation.
Documented delivery scope
The historical project account describes offline leaf-image capture and diagnosis after reconnection. The dataset and crop coverage below describe scope, not diagnostic accuracy or causal agronomic impact.
Crop-specific playbooks and the OpenAI API conversational layer are recorded as intended extensions. Their production release is not established and they are excluded from delivered advisory scope.
Judge the path around the prediction
For an AI product like this I would examine the point of action: can someone collect usable evidence, recover from a failed connection, understand the answer’s limits and take an appropriate next step? Those product decisions belong beside the model output, rather than follow it as interface polish.
What the dataset establishes
The reported dataset covered more than 500,000 images across fourteen crops. That demonstrates the described data scope, not class balance, held-out performance or suitability for every image quality and field condition. The source does not contain a publishable requirement example from the agronomy review.
Keep the advisory maturity explicit
Crop-specific field playbooks and an OpenAI API-powered conversational layer were intended to connect the diagnosis to practical guidance. The records do not establish their production release. They are not presented as a delivered assistant or a measured reduction in support work.
Low-end phones and inference efficiency were delivery constraints alongside connectivity. The account does not include a publishable retry/recovery trace or a defined low-confidence threshold, so those details remain proposed validation questions rather than historical controls.
Current advisory context · October 2026
GrowStac’s October 2026 public page advertises AI/ML advisory as an Upaj checklist with farm-data and agronomy support. This post-tenure proposition supplies market context, not historical proof of the intended conversational layer’s release.
How I would validate this today — proposed
I would evaluate three questions separately: diagnosis on held-out crop examples; capture/reconnection on target devices; and safe, useful advice reviewed by agronomists. The full eight-scenario table is a proposed follow-up plan, not executed validation or an accepted threshold set.
Scroll across to see every column.
| Scenario | Expected behaviour | Available evidence | Status |
|---|---|---|---|
| Representative labelled images | Evaluate diagnosis against held-out agronomist-labelled examples across crop conditions; report errors and per-crop precision/recall. | Existing: source-reported 500K+ image dataset and agronomist labelling/review loop. No held-out examples or model-evaluation results are attached. Proposed: representative labelled evaluation set separated from training data. | Proposed — not executed |
| Crop coverage | Check each of the fourteen reported crops and disclose uneven condition/class coverage rather than treating the total dataset as uniform coverage. | Existing: reported fourteen-crop scope; a per-crop/class evaluation breakdown is unavailable. Proposed: crop/condition coverage inventory and stratified results. | Proposed — not executed |
| Poor image quality | Detect or safely handle blur, poor light and unusable framing; request a better capture or defer advice instead of confident unsupported diagnosis. | Existing: capture-to-advice workflow and low-end-device constraints. No difficult-image evaluation results are attached. Proposed: labelled poor-quality capture cases and recovery checks. | Proposed — not executed |
| Expert agreement | Compare model outputs and advisory with agronomist judgements, recording disagreements and review decisions. | Existing: reported agronomist review loop and crop-specific field playbooks; measured expert-agreement results are unavailable. Proposed: independently reviewed examples with agreement, correction and error analysis. | Proposed — not executed |
| Uncertainty handling | Make uncertainty understandable and abstain or escalate when evidence cannot support a useful answer. | Existing: product account identifies uncertainty and advice as delivery questions. No calibrated-uncertainty benchmark is attached. Proposed: ambiguous, unfamiliar and conflicting cases with abstention and reviewer checks. | Proposed — not executed |
| Field usability and reconnection | Preserve offline evidence capture, recover after reconnection and keep response time and guidance usable on low-end devices. | Existing: described offline capture/reconnection flow and inference-efficiency constraints. Proposed: device, network, successful-capture/recovery and response-time observations. | Proposed — not executed |
| Advisory safety and usefulness | Return correct, useful crop-specific next actions; escalate unsafe or unsupported guidance. | Existing: crop playbooks and the described OpenAI API conversational layer. No measured advisory-quality report is attached. Proposed: expert-reviewed playbooks and realistic farmer questions. | Proposed — not executed |
| Adoption and agronomic impact | Define a user cohort, baseline and follow-up period; distinguish adoption/advice use from causal crop outcomes. | Existing: reported dataset/delivery scope only; causal yield and support-deflection evaluations are absent. Proposed: agreed adoption measures and a separately designed agronomic impact study. | Proposed — not executed |
The test sets, baselines and release thresholds must be agreed before execution. This table records proposed validation work.
Notes and sources
- Historical project/career account supports the dataset, crop coverage and capture/reconnection journey; these are self-reported. Diagnostic accuracy and causal yield or support-deflection outcomes are not established.
- October 2026 GrowStac captures describe a current advisory proposition. The evaluation table is an authored proposal, not a report of tests executed during historical delivery.
Upaj portfolio public reference · Reference pages 1–6 · 13 October 2026 · Post-tenure public product capture
Proposed validation plan · Authored portfolio plan; not executed results