Goji

About

Goji (www.goji.pet) is a dog walking management platform built for agencies that coordinate multiple walkers across a city. At its core is an LLM-powered scheduling engine that automatically matches walkers to walks by reasoning over availability, geography, pet history, walker capabilities, and free-text manager notes.

The system replaces hours of manual scheduling with intelligent suggestions that managers can accept, adjust, or override — keeping humans in the loop while eliminating the cognitive load of juggling dozens of variables across a daily schedule.

Approach

The matching engine uses a three-layer architecture: a deterministic hard filter narrows candidates by availability and coverage area, an LLM global optimizer reasons across all remaining constraints simultaneously, and a heuristic scoring fallback activates automatically if the AI layer underperforms. A safety kill switch triggers at a 40% acceptance rate drop.

AI output is validated against strict schemas. All suggestions include reasoning traces so managers understand why a match was proposed. The system is provider-agnostic — currently using Claude Sonnet for assignments and Haiku for summaries — so the AI layer can be swapped without touching product logic.

Key Decisions

• Provider-agnostic AI layer — abstracted LLM calls behind a clean interface so the product isn't locked to any single provider.
• Deterministic fallback — if the LLM misbehaves, the system degrades gracefully to rule-based scoring rather than failing.
• Free-text constraint parsing — managers can type natural language scheduling rules ("don't send Jake to the Upper West Side on Tuesdays") and the LLM incorporates them.
• Trust framework — identified three critical risks (hallucinated availability, cross-tenant data leakage, demographic bias) and built mitigations for each before launch.
• AI cost modeling — projected AI COGS at 2–6% of revenue, validated that unit economics work at scale.