I build AI systems that earn their keep.

A 6-agent orchestrator that answers phone calls and WhatsApp messages, qualifies the lead, and books the appointment - no human in the loop.

A query rewriter and router run in parallel, dispatching to a knowledge agent, a qualifier agent, or an escalation agent depending on intent, with a confidence-gated RAG layer that never answers past what's actually retrieved. A sixth agent extracts structured lead fields (name, phone, budget) alongside the reply. Twilio media streams into Deepgram STT and back out through TTS in under two seconds round trip. Multi-tenant with row-level security in Postgres, a lead-scoring dashboard, and a vertical-config system so one orchestrator serves different business domains without forking code.

195K+ vector search across 4,491 tax documents and 2,041 case laws, returning cited answers in seconds instead of hours of manual lookup.

Multi-tenant document RAG with per-firm data isolation, freemium billing, and an admin panel for document management. Currently offline for maintenance - this is real production experience at real scale, not a live demo to try today.

A niche goes in, four platforms get a published video out - script, voice, image, captions, human approval gate, zero cloud AI spend.

Fully self-hosted n8n pipeline: local LLM writes the script, local TTS narrates it, local diffusion model generates the hero image, FFmpeg composes the video, WhisperX burns word-accurate captions, a Telegram approve/reject gate holds it before it posts simultaneously to four platforms. Entirely Docker Compose, no per-token API costs.

An agent that reads a resume, then runs daily: pulls listings from five sources, embeds and re-ranks against a preference profile, and emails a digest with an auto-generated cover letter per application.

Built for myself in one session because I was tired of job boards. Five-source ingestion, vector re-ranking with sentence-transformers, LLM-generated per-application cover letters, all on a Celery/Redis job queue.