Senior Full-Stack & AI Engineer building production systems end to end: FastAPI and Go backends, event-driven workers, React and Next.js front ends, Kubernetes on AWS, and the LLM, RAG and agentic tooling that runs on top of them.
Synthetic dataset generation and LLM-as-judge evaluation platforms, an LLM in a production control loop, agentic workflows with guardrails, and MCP clients and servers for my own tools.
MCP both ways, in Go and TypeScriptI’m Cristian Gîrlea, an engineer and technical lead with 13+ years across fintech, AI, cybersecurity, healthtech and enterprise platforms. Lately that means LLM evaluation platforms, agentic engineering workflows and MCP servers for my own products, next to the ingestion pipelines and cloud infrastructure that keep them honest. My focus is clear: production-grade services, calm incident response, and teams that understand the systems they operate.
Own full vertical slices of a brand-protection ingestion platform on AWS EKS: Go consumers, TypeScript crawlers, KEDA-scaled queues and staged customer rollouts. Shipped two advert-intelligence sources end to end, put an LLM in the proxy-control loop, built a crawler canary system for 20+ platforms and a library of 30+ guardrailed agent skills.
Sole engineer and architect of the synthetic data-generation service inside an LLM evaluation platform: FastAPI and Flask, Hatchet workers, idempotent async batches persisted as MLflow evaluation datasets, LLM-as-judge validation, organization-scoped access control, and a Docker split that cut incremental rebuilds from ~40 to under 10 minutes.
Modernized projection and portfolio-simulation services across Go, FastAPI and NestJS with Protobuf APIs, added an LLM-backed explanation service, and built AWS serverless marketplace services for a lending platform: ~40% faster APIs and 15+ services on EKS and AKS with OpenTofu and GitHub Actions.
Improved operational efficiency 5x with actionable analytics, cut recurring operational workload by 80%, and improved system performance by 50% through workflow and query optimization.
Led the modernization of a crisis-region platform, improving performance 10× and reducing dashboard load time from 20 to 2 seconds, and integrated Azure AD single sign-on (OIDC) for staff access.
Windows remote desktop · Early alpha
Your computers. Within reach.
I’m building a portable Windows remote-desktop app in Rust, bringing screen streaming, system audio and keyboard/mouse control together over an encrypted QUIC connection.
From native screen capture and media encoding to configurable clipboard sharing and automated release packaging, it’s a hands-on systems project built around connecting my own computers.
Source available. New development is free for personal, non-commercial use.
More projects · github.com/cristiangirlea
SDLC kit for Claude Code and Codex: spec, plan, implement, review, verify, ship and learn as evidence-gated stages, with 15 skills, 11 scoped agent roles, a local issue tracker and hooks that block destructive commands.
Personal agent harness with its own Anthropic API agent loop and Claude Code, Codex, Gemini and opencode backends. Speaks MCP both ways over stdio, records every run, and batches tasks across parallel git worktrees with gates and retries.
Job discovery and application tracker over six public job APIs, with deterministic document generation and an MCP server on the official Go SDK that exposes matching and tracking to agent hosts under the user’s own token.
Multi-country invoicing SaaS with an EN 16931 model, UBL 2.1 generation and e-Factura, MTD and Peppol adapters, plus a token-authenticated, organization-scoped MCP server so assistants can read the books and draft invoices.
Finds funded open-source issues across GitHub, GitLab and Forgejo and runs coding agents on them in one hardened container per repository, publishing only when the project’s own tests pass.
Local control plane for delegating work to coding agents: isolated worktrees, per-stage model routing, policy limits and a human approval gate before anything merges.
Typed PHP client and Python task service for bounded AI work: async jobs, per-task process isolation, ONNX Runtime cross-encoder reranking and fault-injection tests.
Seven patches to math, math/big, math/rand and crypto/tls, two merged after Go team review and five under review, plus open pull requests to PostHog’s Go SDK and Canonical’s Juju operator framework.
Benchmark harness for local LLMs served by Ollama, with a web UI, CLI, Redis-backed worker queue and PostgreSQL.
Hands-on engineering backed by structured troubleshooting, customer communication, documentation and knowledge sharing.
Go · Python · TypeScript · JavaScript · SQL · PHP
FastAPI · Go services · Node.js · NestJS · REST · gRPC / Protobuf · webhooks · async jobs
LLM integration · RAG & reranking · LLM-as-judge evaluation · MLflow · MCP clients & servers · agent skills, roles & hooks
React · Next.js · Vue · Angular · Tailwind CSS
Kubernetes · Docker · AWS EKS · Azure AKS · Google Cloud Platform · Terraform / OpenTofu · KEDA · Flux
GitHub Actions · GitLab CI · Azure DevOps · canary & end-to-end testing · OpenTelemetry · Prometheus · Grafana · Sentry
PostgreSQL · MySQL · MongoDB · Redis · Elasticsearch · Kafka · RabbitMQ · Beanstalkd · Hatchet
University of Craiova