I build systems that work while you sleep.

SYSTEM / 01 — PROSPECT

Every morning, Prospect scans newspapers with OCR, extracts job postings with LLMs, matches them against persona-scoped resumes with semantic search, and delivers ranked alerts to Telegram. A multi-agent system — scanner, context engine, generation runtime, tracker — built and run on my own hardware.

No job boards, no third-party crawlers. Prospect reads the same newspapers a human would — then does the part humans skip. It OCRs every page, extracts each posting with an LLM, embeds and scores it against persona-scoped resumes, and writes the ranked shortlist to my phone. Eleven stages, running unattended on hardware I own.

SCAN → EXTRACT → EMBED → MATCH → GENERATE → DELIVER

SYSTEM / 02 — TRAVEL PLANNER AGENT

An agentic AI travel planner with a deterministic memory layer, a custom model router, and an async circuit breaker — when cloud APIs degrade, it falls back to local Ollama inference and keeps streaming. Deployed on a k3s multi-node cluster with full Prometheus/Grafana observability.

Cloud APIs degrade. Rate limits hit. This agent is built for that moment: a deterministic memory layer so retries never repeat work, a router that picks the right model per task, and an async circuit breaker that swaps to local Ollama inference mid-stream — the response never stops. Deployed on a k3s multi-node cluster with Prometheus and Grafana watching every request.

Where the systems shipped

Stack

Verifiable elsewhere