I was deep in my AI learning journey when I started hearing the same complaints from budtenders over and over — having to field the same questions endlessly, staff stretched thin during peak hours, and customers grumbling about the wait. I realized this was the perfect opportunity to apply AI concepts to a real-life problem. One where I could test and deploy in the real world. What if the knowledge of your best budtender was always available and never burned out?
That was the seed of the AI Budtender.
The cannabis industry is a fascinating and unforgiving environment for AI. It’s heavily regulated, privacy-sensitive, and operating in a legal gray zone at the federal level. Every design decision has a compliance dimension. That constraint, it turns out, makes for better software.
What It Does The AI Budtender is a customer-facing assistant that delivers personalized cannabis product recommendations based on a live inventory of 2,000+ SKUs and individual purchase history. It understands natural language queries, respects California compliance requirements, and responds in under 10 seconds — because that’s the threshold between helpful and frustrating in a retail environment.
The Stack
- Frontend: Next.js / Tailwind CSS, deployed on Vercel
- Backend: FastAPI, deployed on Railway
- Database: Supabase PostgreSQL with pgvector for semantic search
- Embeddings: OpenAI
- LLM Access: OpenRouter (multi-provider)
- Compliance: NeMo Guardrails
- Analytics: PostHog
Hard-Won Insights
- Small open-source models lack the conversational coherence needed for a 2,000+ product catalog
- Prompt-based compliance isn’t enough — California regulators need proof of technical controls, not policy statements
- Inventory awareness isn’t a feature, it’s the foundation of customer trust
Status: Currently in testing with budtenders, validating that recommendations are grounded in actual inventory. A model that confidently suggests products you don’t carry isn’t helpful to the store — it’s sending customers elsewhere. The AI Budtender should be the dispensary’s best salesperson, not an accidental advertisement for the competition.
Future iterations will go further — surfacing active brand promotions, in-store discounts, and featured products at the right moment in the conversation, turning every customer interaction into an opportunity the store controls.