The Support Agent
RAG-based support that refuses to guess
LangGraph agentCorrective RAGGuardrails
2024 - now
Work
Systems that survived production.
Three shelves: the flagship AI platform I build end to end, the products and client work that shipped along the way, and what I build on my own time.
01 — Flagship
Harrir حرير — four systems, every layer mine.
An AI-native fashion platform for women in Iraq — two agents that speak four registers of Arabic, search built for dialect, and a 16GB GPU serving all of it. Each system has its own scars and its own write-up.
Harrir — the flywheel
How the models get better while I sleep.
- 01
Ship
Agents answer real customers in four registers of Arabic and English. Every turn is captured — the conversation is the dataset.
- 02
Judge
A 32B model grades each new turn against corrected hindsight — later turns judged as if earlier mistakes were already fixed.
- 03
Correct
Humans step in only where the judge flags trouble. One reviewed correction is worth more than a thousand scraped examples.
- 04
Amplify
Each correction becomes eight varied synthetic conversations ending in the exact corrected tool call — spread across every register.
- 05
Promote
Retrain the LoRA, replay champion vs challenger on the live serving path, and hot-swap the adapter. No restart. No regression. Repeat.
02 — Products & client work
Shipped, live, and paying rent.
Websites, platforms, and a Play Store app — the product engineering that keeps the AI work grounded in real users and real deadlines.
03 — Personal projects
Built on my own time.
Where I try the ideas nobody commissioned, and find out what they cost before recommending them to anyone else.