MyLawn Brain
In Build · Function: Consumer / Digital · Where it runs: Cloud Run + Cloud SQL
The first initiative to consolidate Scotts's 150 years of lawn-care expertise into one auditable, always-consistent intelligence layer that every consumer channel can call.
Problem
Scotts has the deepest lawn care expertise in the industry — codified in product labels (the regulatory law for when, how, and where each product can be applied), knowledge articles, training materials, and decades of R&D guidance. The material is not missing — it's scattered. PDFs, spreadsheets, internal docs, no single machine-readable form.
As a result, different consumer channels can give different advice. The app says one thing, the website another, customer service a third. No single structured source is feeding every channel.
What it does
- Extracts structured data from every active product label and supporting article using agentic AI — approved grass types, prohibited grasses, listed weeds, diseases, pests, application rates, temperature constraints, seeding restrictions, frequency limits.
- Structures the legacy: 26 grass types, 152 weed species, 37 diseases, 31 pests, 14 regions, 11 seasons. 100+ products organized into 48 functional subcategories by the job they do (not just the shelf they sit on).
- Reasons: 126 rules evaluate against a 48-field user state vector — grass type, region, photo analysis (weeds, disease, bare spots), weather (GDD, soil temp, humidity), and full action history. 14 hard safety rules + 10 soft advisory rules, codified straight from the label.
- Serves every consumer touchpoint with the same recommendation: MyLawn App, the website, Sierra Chat, CRM/email — one brain, many channels.
Every recommendation is traceable — you can see exactly which rule fired and why.
Who owns it
- Business owner: Consumer Digital / I&TC
- CoE AI Solution Lead: [TBD]
- Tech stack: Python rule engine + Cloud SQL Postgres + Cloud Run, agentic AI extraction pipeline, React admin UI (brain-admin)
Where it runs
- Backend / rule engine:
brain-services— Cloud Run, Cloud SQL (Postgres), private behind a Cloud Function BFF - Admin UI:
brain-admin— Vite + React, for managing the rule corpus and previewing recommendations - GCP project:
mkt-dev-0-0
Data it touches
| Data source | Classification | Access path |
|---|---|---|
| Product labels (PDF + structured) | Public + Regulated | Agentic extraction pipeline → structured Postgres |
| Knowledge articles, training material | Internal | One-time ETL into structured form |
| Active product catalog | Internal | Synced from source systems |
| User profile (grass type, region, action history) | Consumer PII | Read via authenticated API, no PII in prompts |
| Weather (GDD, soil temp, humidity) | Public | External API |
Governance status
- AI Governance Committee reviewed: Pending formal review — extraction pipeline and rule engine treated as agentic AI under FY27 governance pattern
- Human-in-the-loop required for: Rule changes promoted to production (validated via simulation against historical journeys)
- Vendor / model risk noted: Yes — agentic extraction uses third-party model; outputs reviewed before promoting to rule corpus
Outputs
- Products — tailored recommendations from 100+ SKUs across 10 categories
- Actions — engagement tracking; what the consumer did, when, what to do next
- Lawn Score — lawn-condition assessment from photo analysis + history
- Guardrails — 14 hard constraints (e.g. "never seed within 16 weeks of non-seed-safe pre-emergent") + 10 soft advisories prevent bad advice across every channel
- Alerts — proactive outreach based on GDD, weather, and timing windows
- Simulation — validate rule changes before promoting to production
Patterns reused
- Agentic extraction → structured-data pattern is reusable for other label-and-document corpora (e.g. safety data sheets, OG&L technical bulletins)
- Rule-engine + state-vector evaluation pattern is reusable for any recommendation surface that needs traceability and guardrails
What's next
- Expanding rule coverage to lime/sulfur, aeration, dethatching
- Weed-specific product matching (photo identifies dandelion → recommend products that list dandelion on the label)
- Multi-year journey simulation to validate recommendation quality over time
- Production rollout across all four consumer channels (app, web, Sierra, CRM)
Changelog
- 2026-05-21 — Initial entry on the hub.