Arif D.

Full-stack AI developer and backend engineer. I build AI systems that survive contact with real users: AI agents, RAG pipelines, LLM apps and voice AI, plus the full-stack SaaS products around them. Python (FastAPI) and Go (Golang) backends, TypeScript with Node.js, Next.js and React frontends, PostgreSQL and AWS underneath. Based in Berlin, 10 years in backend and full-stack development, working with teams across Europe and the US. German invoicing, EU data residency when you need it. What I build: - AI agents and AI automation: agents that use real tools, call your APIs, retry when something fails and hand off to a human when being wrong is expensive. - RAG and LLM apps, including AI chatbots that answer from your own documents: retrieval, evals, structured output and guardrails, so answers are correct and not just confident. OpenAI, Claude and open models. - Voice AI: inbound and outbound voice agents with live transcription, intent routing and escalation to a human. - Full-stack SaaS: Next.js and React frontends on Python, Go or Node.js backends, from MVP to production. - Backend development and API integration: REST and WebSocket APIs, CRM integrations, Stripe payments, OAuth, Google Workspace. - DevOps: Docker, Kubernetes, Terraform and CI/CD on AWS and GCP. A few things I shipped recently: AI voice receptionist backend for a US SaaS. Inbound call handling, live transcription, intent routing, and escalation to a human when confidence drops. Three weeks from empty repo to production. The client rated it 5.0 and came back for four more projects. German legal RAG chatbot. Vector search over statutes and case law, plus a fine-tuned LLM for domain chat. I built an eval set before touching anything, which is how I know retrieval went from 3 correct out of 12 test questions to 10 out of 12 rather than just feeling faster. Most of that gain came from query rewriting, not from a bigger model. Call tracking SaaS backend. Seven milestones over four months: user lifecycle, server-side search with proper indexing, CRM integrations, OAuth hardening, a recording pipeline, audit logging, staging with auto-deploy. Every milestone 5.0. EU compliance engine in Go. Driver hours and rest periods under Regulation 561/2006, encoded as a rules engine with a test per legal clause. Document AI pipeline. 100k+ real documents (PDFs, scans, Office files, CAD exports, emails) at a 99% success rate, with OCR and LLM extraction in fallback and vector search feeding a RAG platform. Background: - Co-founder and CTO of an AI productivity SaaS for four and a half years. Raised $1M pre-seed and ran engineering. Split a Laravel monolith into 8 Go microservices plus Python AI agent services: p95 latency down 55% at about 3k requests per second. Our AI agents handled about 400 tasks a day with under 2% human escalation. - Fintech backend at a crypto exchange with 1M+ users: scaled a Python (Django) trading engine to 30,000 concurrent users and 15 million order events a day. - Full-stack and backend contracts: Go booking and payment microservices with REST and WebSocket APIs and Stripe subscriptions for a marketplace, and CI/CD for 20+ microservices that cut deploys from 30 to 8 minutes. - Open source: GoVisual, a Go HTTP debugger (about 680 GitHub stars, awesome-go, Hacker News front page), and fastapi-radar, a FastAPI monitoring dashboard (about 400 stars). Speaker at the GDG Berlin Golang meetup. Where I'm strongest: LLM applications that have to be correct. Retrieval, evals, structured output, guardrails. I fine-tuned a German legal LLM and shipped the earlier checkpoint because it scored better on hard negatives, while the newer one looked better on easy examples and fell apart on the ones that mattered. AI agents and AI automation that genuinely use tools and handle failure, not prompt chains. Retries, escalation paths, and a human in the loop wherever being wrong is expensive. Python, Go and TypeScript backends and full-stack apps built to be handed over. Tests that mean something, OpenAPI docs that match the code, migrations that run cleanly from empty, and a deploy someone else can trigger without calling me. Stack: Python, FastAPI, Django, Go (Golang), TypeScript, Node.js, Next.js, React, PostgreSQL with pgvector, Qdrant, Redis, OpenAI and Anthropic APIs, LangChain where it helps and plain code where it doesn't, Docker, Kubernetes, Terraform, AWS, GCP. Where I'm not the right person: foundation model training, data labelling, and anything where the brief is "build me a ChatGPT clone". I reply within a few hours during European business time. Tell me what the system has to get right and what happens when it gets it wrong; that tells me more than a feature list does. If it isn't a fit I'll say so in the first message instead of billing you to find out.