Prompt Saz
Trilingual AI prompt gallery & semantic search engine
- Role
- Full-Stack Architect & AI Integrator
- Status
- Live
- Year
- 2026
- Duration
- 1 Year (October 2025 – October 2026)
- Stack
- Next.js
- Supabase
- pgvector
- Gemini API
- PostgreSQL
- Skills
- Semantic Search
- Vector Databases (pgvector)
- LLM/API Integration
- Trilingual Localization
- Links
- Live demo
Results
HNSW
Vector Indexing
Problem
Users lacked a centralized, localized platform in Kurdish, Arabic, and English to discover, test, and analyze effective AI prompts using advanced semantic matching rather than basic keyword search.
What I built
- Architected a Next.js and Supabase infrastructure with comprehensive trilingual routing and localization.
- Engineered a hybrid semantic search engine utilizing pgvector, HNSW indexing, and Gemini embeddings to accurately discover relevant prompts.
- Developed an admin analytics dashboard and collaborated on a formal academic usability paper to evaluate the platform's user interface.
Gallery
Key decisions
Native pgvector vs Third-Party Search
Why
Kept the vector database native to Supabase to reduce infrastructure complexity, using Gemini embeddings for high-quality multilingual semantic understanding.
Trade-off
Required deeper knowledge of vector math and manual database tuning (HNSW index optimization) compared to using a managed service like Algolia.
Technical notes
HNSW Indexing Optimization
Implemented Hierarchical Navigable Small World (HNSW) indexing directly in PostgreSQL to achieve sub-millisecond semantic similarity matching on large datasets of prompts.
What I would do next
Prompt Saz was my most architecturally complex project, bridging full-stack web development with machine learning integration. Co-authoring a research paper on its usability was a significant milestone in my academic and professional journey.
Need something like this?
Available for full-time frontend/full-stack engineering positions, contract roles, and impactful civic tech initiatives.