startup
Artificial Mufti
AI-powered comprehensive Islamic knowledge and guidance platform
The Problem
Traditional search engines return surface-level or contradictory Islamic rulings — the domain requires nuanced, source-verified answers referencing classical jurisprudence, which keyword search cannot provide.
The Solution
Built a two-stage RAG pipeline in Python/FastAPI: a retrieval confidence scorer filters low-certainty chunks before LLM inference, and a post-generation validation layer cross-references responses against the source corpus using cosine similarity thresholds — reducing hallucinations without increasing latency.
My Role
AI Solutions Architect & Full-Stack Developer
Timeline
6 months
Product
Architecture
Frontend
Next.js (App Router)
Backend
Python / FastAPI
Realtime
Server-Sent Events (SSE)
Infrastructure
AWS ECS & Vercel
Tech Stack
Impact
Engineered a two-stage prompt-validation pipeline (retrieval confidence scoring + post-generation cross-reference check) that reduced hallucination rate from 18% to 9.9% on an internal evaluation benchmark — critical for a domain where factual precision is non-negotiable.
Achieved <800ms average response time for complex semantic vector searches across a 500K+ document corpus in Pinecone by implementing hybrid dense-sparse retrieval and async SSE streaming — users see tokens within 400ms.
