# Devrazzi

> Devrazzi is a content product that leaves the signal/noise problem in developer news not to a human editor but to a layered model pipeline. It collects sources, runs a first pass locally with Ollama, hands off whatever clears the bar to a step-specific model among DeepSeek/Gemini/OpenAI/Claude, and produces and publishes a TR/EN synthesis end to end. Live on web and iOS; full article extraction isn't there yet, and that's kept open as a known limit.

- Role: Founder & Developer
- Class: Side Publications
- Status: Ongoing
- Started: 2026-09-23
- Technologies: TypeScript, Node.js, Prisma, PostgreSQL, Ollama, DeepSeek, Gemini, OpenAI, Claude, Fastify, Astro
- Skills: Pipeline Engineering, LLM Orchestration, Prompt Engineering, Content Automation, API Integration, Data Pipeline Design
- Website: https://devrazzi.com
- App Store: https://apps.apple.com/us/app/devrazzi-curated-tech-news/id6790693413
- Source: https://muhammetsafak.com/portfolio/devrazzi/
- Language: en-US
- Author: Muhammet Şafak

---
Devrazzi is a content product that leaves the signal/noise problem in developer news not to a human editor but to a layered model pipeline: it collects sources, scores them, filters out whatever falls below the bar, and produces and publishes what's left as TR and EN synthesis. The first pass runs locally — a model running on Ollama scores every record on originality, technical depth, pragmatism and ecosystem impact, and a record whose average lands below 7 drops right there. Whatever clears the bar is handed off to a step-specific model among the DeepSeek, Gemini, OpenAI and Claude APIs; the model's output is bound to a strict JSON contract, and a response that doesn't hold the contract is rejected.

The pipeline is live on the web at [devrazzi.com](https://devrazzi.com) and in the iOS app; I built the state machine (RAW → EVALUATING → APPROVED/REJECTED → PUBLISHED) on modular TypeScript packages, on top of Prisma and PostgreSQL's JSONB. I'm leaving a known limit openly on the table: there's no full article extraction yet, so sources that feed "link + two sentences" — Hacker News and Reddit among them — get filtered out for insufficient content. The next job is closing that blind spot and pulling the cost/quality measurement per tier out of gut feeling.