AgriScout

Tagline: Snap a leaf. Get a diagnosis and a next step.

Comment

Inspiration

Smallholder farmers lose a large share of their harvest to diseases and pests that are caught too late. Extension officers are few and far away, but almost every farmer has a phone. We wanted a system that does more than label a photo: it looks at the evidence, decides what it needs next, and brings in a human when the stakes are high.

What it does⁸

AgriScout turns a leaf photo into a diagnosis and a next step, and lets the visual evidence drive the workflow: Blurry or dark photo → asks for a better retake, with specific guidance Ambiguous result → runs a second pass (zoomed crop or alternate model) High confidence → writes a plain-language advisory in English or Swahili Severe or uncertain case → escalates to a human extension officer before any advice is sent

How we built it

1.Quality gate (OpenCV 5): blur is scored with the variance of the Laplacian, and a photo is accepted only if plus an exposure check on the histogram.

  1. Vision pipeline (OpenCV 5): leaf segmentation, lesion detection from color and texture, and classification with the dnn module.
  2. Agent (Amazon Bedrock): a tool-using agent chooses request_retake, reanalyze_crop, write_advisory or escalate_to_human from the vision output.
  3. Escalation rule: the agent acts alone only when otherwise a human approves first.
  4. AWS (Graviton/arm64): S3 for images, Lambda/ECS for OpenCV workers, DynamoDB for the case log, CloudWatch for monitoring, IAM least-privilege, all deployed with infrastructure as code.
  5. COOL benchmark: the same pipeline on x86 vs Graviton, measuring latency, throughput and cost per 1,000 images. ## Challenges we ran into Field photos are far noisier than benchmark images, so the quality gate became essential Deciding when the agent should act alone and when it should hand off Keeping inference fast and cheap on CPU-only arm64 ## Accomplishments that we're proud of A working end-to-end loop: photo → retake → diagnosis → escalation Published Graviton vs x86 results: [X%] lower cost per 1,000 images [MEASURE] One-command reproducible AWS deployment ## What we learned Vision confidence should steer the whole workflow. Reacting to image quality improved results more than a larger model did. [EDIT TO MATCH YOUR REAL FINDINGS] Results: accuracy [X%], macro-F1 [X], task success [X%], escalation rate [X%]. ## What's next for AgriScout Offline-first mobile capture, more crops and local languages, and validation with extension services.

Built With

Share this project:

Updates

Submission history