Dashgin Khudiyev
software engineer · ai systems · self-hosted
Python & Django backends, LLM products built with data privacy in mind, payment-critical systems, React & React Native up front, running on infrastructure I operate myself.
Your coding agent draws the screens over MCP, and Loupe renders each one in a real browser, measures it, and checks it against your design system. The options sit side by side on a board: you pick one and say why, and the agent builds from the version you picked, not a description of it. Slides and launch posts go on the same board.
A fully local macOS companion that lives in the notch. Hold the globe key, point at something and say what you want, and Puck does it on the Mac in front of you. A grammar decides what you meant, a gate checks it, and a local model only writes the words it says back. It asks before anything that can't be undone. Puck Remote turns an iPhone into a second talk key.
A browser MCP server for AI agents: in a benchmark with Claude it took 7 turns where Playwright MCP took 13, at half the cost. The page comes back as a numbered list of what the agent can click, and eight tools cover the rest in under 1,000 tokens. Building it turned up a Playwright bug; I sent the fix upstream.
A car marketplace intelligence platform. I scrape turbo.az listings, price every car against real comparables, and alert users the moment something appears well below market value. Django/Ninja backend, an Expo/React Native mobile app, and a server-rendered web app share one API.
An education platform for teachers and students. I built the React web app, the Django backend, and the React Native mobile app.
A horizontally-scaled control plane for marketplace seller automation. A single Django image runs four roles (web, worker, scheduler, and a per-shop agent), where a central control plane orchestrates a fleet of per-shop agent VMs. Each agent reaches a shared PostgreSQL (through PgBouncer) over a private WireGuard mesh (NetBird), so shops scale out one VM each while all state stays central. A push to main auto-deploys the control plane and rebuilds every agent in the fleet.
In the Contractors product of a global hiring & payroll platform, I built the first version of RemoHire, a multi-agent AI interviewer (Pydantic AI) that gathers job requirements from clients, with separate agents for asking questions, extracting requirements, and keeping the conversation on track. I also integrated Airwallex for real-time payouts and rewrote payment-file processing to handle 1,000+ payments per batch.
An AI-powered brand asset hub. It centralizes logos, templates, videos and guidelines with auto-tagging and natural-language search, version control and approval workflows, and multi-tenant sharing for agencies. 50+ file types, connectors for Drive/Dropbox/S3.
A real-estate platform that scrapes listings, predicts prices with an ML model, and plots everything on a map.
Part of the team building a supply-chain management SaaS on a microservices architecture with CQRS and Event Sourcing. I designed and implemented services, database schemas, deployment, and monitoring.
Backend for a hiring platform connecting IT companies and candidates: job posting, applications, and subscription plans with recurring payments.