Kaloyan Yordanov

I build and deploy AI systems that solve real problems, from multi-agent architectures to systems that real businesses run on.

Selected work

3 of 4 live in production
  1. Live

    LLM Gateway

    A production gateway that sits in front of LLM providers. Built in Java.

    One self-hosted service that gives every app a single, controlled way to reach an LLM. It's now multi-tenant and observable. Each client gets its own authentication, rate limits, and a hard budget cap, all managed from a built-in admin panel that creates, edits, and disables clients and sets their budgets. Once a client hits its cap, it's rejected before the provider is ever called, on top of response caching, cost accounting, automatic failover between providers, per-model routing, and live streaming. Request-latency percentiles (p50/p95/p99) surface on the dashboard. The core engineering principle is that correctness is measured, not asserted: an un-bypassable build gate fails unless test coverage, mutation testing (which proves the tests actually catch bugs), and static analysis all pass. Built with AI-assisted development held to that gate the whole way. The public demo runs a spend-proof stub provider, so it costs nothing to explore.

    • Java 25
    • Spring Boot
    • PostgreSQL
    • Docker
    • Resilience4j
    • Micrometer
    • Flyway
    • React
    • Cloudflare Tunnel
    • Hetzner VPS
    Live demo GitHub
  2. Live

    Agent Central

    A visual multi-agent observability system.

    Orchestrates a team of coordinated AI agents (a job scout, analyst, deal hunter, and secretary) with a live interface where each agent's real backend state drives its animation, so the system's behavior is observable rather than a black box.

    • Python
    • FastAPI
    • PostgreSQL
    • React
    • Phaser
    • Anthropic API (Claude)
    • Cloudflare Tunnel
    • systemd
    • Hetzner VPS
    Live demo GitHub
  3. Live

    Hotel Reservation System

    A full-stack booking system, deployed and running a real family hotel.

    Real users depend on it daily. Its core engineering decision: double-booking is made structurally impossible at the database level, using a Postgres exclusion constraint over date ranges, rather than application-level checks that can race under concurrent writes. Built for correctness and real-world reliability, with database-enforced guarantees and nightly backups, not a demo.

    • FastAPI
    • PostgreSQL
    • React
    • Vite
    • Alembic
    • Cloudflare Tunnel
    • Hetzner VPS
    Live demo GitHub
  4. Coming soon

    Kaloyan's Assistant

    A real-time voice agent that answers the phone and talks about my work.

    Built on a native audio-to-audio model, it handles live phone calls with natural interruption (barge-in), holding a real spoken conversation rather than a rigid menu. Callers can hear it describe these very projects.

    • Twilio
    • OpenAI Realtime API (gpt-4o-realtime)
    • Python
    • FastAPI
    • WebSockets
    Coming soon, callable phone demo

Contact