{
  "id": 17081609,
  "title": "Why We Stopped Building Just Another Plant Identifier (And What We Learned)",
  "url": "https://urgent.news/2026/10/29/why-we-stopped-building-just-another-plant-identifier-and-what-we",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-10-29T15:37:17.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/plantfinder/why-we-stopped-building-just-another-plant-identifier-and-what-we-learned-23ig"
  },
  "original_language": "en",
  "account": "When the team first began developing Plant Finder, they believed the purpose was straightforward: create an AI model capable of identifying as many plant species as possible from a single photo. They envisioned users snapping a picture of a leaf and receiving an accurate species name with high confidence, typically 99%. However, after extensive real-world trials involving thousands of gardeners, students, and plant enthusiasts, the team quickly discovered an important truth about how people truly utilize plant identification applications. The critical insight: People don’t merely seek a botanical name. Instead, they turn to the app when confronted with an issue: yellowing leaves, unsightly brown spots, or concerns about soil conditions. A botanical name alone cannot remedy these problems; users require guidance rather than merely an identifier.\n\nThe first challenge encountered was the \"Yellow Leaf Problem.\" In agronomy, various issues manifest as yellow leaves, making diagnosis difficult from just visual inspection. Overwatering, underwatering, insufficient nitrogen, and high soil pH all lead to yellow leaves. Consequently, simply identifying the plant based on color would result in incorrect advice that could harm the plant. To address this, the team had to examine the underlying causes of yellowing. In botany, nutrients move differently through a plant. Mobile nutrients, such as Nitrogen and Magnesium, travel upward to support new growth. If the lower leaves turn yellow while the upper leaves remain green, the soil is likely deficient in nutrients or overly acidic. On the other hand, immobile nutrients like Iron and Calcium cannot move once absorbed, and yellowing of the newest top leaves between the veins typically indicates alkaline soil that locks up iron at the roots. By instructing users to examine both the top and bottom of their plants, the app could estimate soil health and pH issues without necessitating the purchase of expensive soil testing kits.\n\nThe second challenge pertained to the computational expense of visual AI. Processing high-resolution leaf photos with large language models consumes significant server resources. As thousands of users upload photos daily, server bills quickly escalate, and users on slow mobile connections experience delays, often abandoning the app due to prolonged waiting times. To mitigate this issue, the team discontinued sending every uploaded photo directly to large language models. Instead, they implemented a rapid visual matching filter on their backend. If a user submits a common leaf or a plant already verified in their botanical database, the system recognizes the visual pattern instantaneously, providing an answer within a second without incurring substantial compute costs. The more complex AI models are then engaged only when faced with unfamiliar or intricate problems.\n\nThe third challenge involved ensuring the app’s security. Initially, many mobile apps made the mistake of executing AI services directly on the phone, leaving private API keys vulnerable to decompilation. Once exposed, unauthorized parties could abuse these keys, resulting in hefty bills. To safeguard against this risk, the team relocated all AI interactions off the phone and secured them behind their own server gateway. The mobile app never directly accesses the AI credentials; it merely sends the photo, verifies the request, and receives diagnostic advice from the secure server. By implementing this approach, the team ensured that sensitive logic and private keys remained protected from potential exploitation.",
  "summary": "When we first started building Plant Finder, we assumed the goal was simple: train an AI model to recognize as many plant species as possible from a single photo. We thought that if a user points their phone at a leaf and the app returns “Monstera deliciosa” with 99% accuracy, the job was done. We were completely wrong. After months of real-world testing with thousands of gardeners, students, and…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}