hi, I'm Adnan Rahmanpoor
I build production AI systems and data products - from evaluation infrastructure, semantic caching, and multi-agent orchestration to a live Dubai real estate intelligence platform processing millions of DLD transactions end to end.
AI SYSTEMS · AGENTIC WORKFLOWS · DATA PRODUCTS · AI-NATIVE DEVELOPMENT
A unified LLM gateway: semantic caching with pgvector cuts response latency by up to 97% on repeated queries, a prompt registry enables weighted A/B rollouts without redeploys, and a background AI-SRE agent diagnoses latency anomalies and pages Telegram with root-cause analysis before a human notices.
A CI/CD pipeline for LLMs. Celery workers run prompt evaluations against golden datasets at scale, an anti-sycophancy LLM-judge grades business accuracy instead of instruction-following, and full OpenTelemetry tracing through Jaeger pinpoints whether a slowdown is Postgres, Redis, or the model API — down to the millisecond.
An independent, DLD-data-driven intelligence platform for Dubai real estate. A DuckDB pipeline unifies historical and daily transaction schemas, computes rental yield, momentum, and liquidity scoring per area, and pre-renders everything to static JSON served through Next.js — no live backend, no cold starts.
Four specialized agents — technical, market, news, sentiment — run concurrently via native asyncio and synthesize into a single hedge-fund-style investment report. Built without LangGraph or CrewAI by design: a custom orchestrator kept the container lightweight and gave full control over failure handling.
I use AI coding agents - Claude Code, ChatGPT, Deepseek and Qwen - as a core part of my development workflow. They accelerate implementation; I own the architecture, trade-offs, debugging, evaluation, and production decisions.
Implementation, boilerplate, first-pass code
Architecture, trade-offs, debugging
4 shipped systems, solo
I started in business and finance, a degree in Management and Finance from Ajman University - rather than computer science. That background shapes how I build: I start with the problem, understand the constraints, and then choose the technology.
Before AI infrastructure, I worked in Dubai real estate market analytics and client-facing property advisory, which is where GetPropIntel came from a need I understood firsthand, not a hypothetical one.
Currently: Dubai, UAE