The models are built. The last mile isn't. I install AI into businesses that run on trucks, permits, and phone calls.
Everything on this page runs in production — agents doing real work with real money and real customers, not demos. I'm an operator, not a vendor: twenty years of P&L ownership, and today I design, code, and self-host AI systems end to end. Most of it I architected and shipped myself; two products are built by developers under my direction.
AI does the persistence and the volume; humans approve what matters; everything is logged.
A Markdown-native workspace with Git as the source of truth, an in-browser editor with an AI coding agent pre-installed for every teammate, and a live operations dashboard — all behind a single MCP gateway (one Cloudflare Worker handling sign-in and SSO, routing internal AI tools like memory, SEO, PPC, and analytics, and audit-logging every tool call for cost attribution). The team works inside it daily.
An always-on agent that builds campaigns, monitors performance, defends against the platforms' auto-applied changes, and proposes fixes with one-tap approval over Telegram. CRM-grounded, governed by explicit written rules, every action logged.
A self-hosted agent exposing dozens of tools across analytics, discovery/SEO, channel management, captions, and bulk reporting — multi-channel, queryable in plain language through an AI client.
Marco builds the sending layer first: secondary domains, SPF/DKIM/DMARC, mailboxes, warm-up, placement tests, per-mailbox caps, and reputation monitoring. The same control plane is designed to scale from one mailbox to a thousand. No campaign launches until its senders clear the deliverability gates.
Then it finds the right moment, not just the right job title. Marco searches across commercial databases, LinkedIn, company sites, job boards, industry directories, and public or first-party data. It looks for signals such as leadership changes, hiring, funding, website intent, and new permits. It scores and deduplicates prospects, researches each company, then writes a short message in the sender's voice. Email and LinkedIn sequences run automatically; Marco triages replies every 15 minutes and flags positive intent in Telegram. Factual claims, pricing, and objections stay behind human approval.
An open database tracking which behind-the-meter power projects at data centers are actually operating. Watchers monitor permits, filings, and public registries; the pipeline archives each source, uses AI to extract and cross-check claims, then holds publication for human approval. Every fact retains source-level provenance and is exposed through CSV, JSON, OpenAPI, and read-only MCP.
A marketplace for ADU construction: complete information architecture, programmatic pages, structured data — plus the full acquisition machine behind it: funnel, voice AI agent answering inbound calls (Vapi), CRM integration, and analytics. Business thesis to production, no handoffs.
Moved recurring vendor tools onto self-hosted open source I selected, deployed, and integrated: Frame.io → FreeFrame, DocuSign → Documenso, Attio/GoHighLevel → Twenty CRM, Calendly → Cal.com. I run the whole layer — Docker, reverse proxy, TLS, backups — plus invoicing automation wired into Mercury.
A concierge agent for general contractors that files and schedules building inspections with LADBS — on top of a compliance engine scoring every contractor on how they actually perform: first-pass rates, violations, permit history. Runs on ~11 million data points covering every LA inspection over 13 years.
An AI that actually watches and listens to a video — footage, audio, pacing, hooks — and grades the edit against two rulebooks at once: the brand's and the mechanics of viral performance. A clear score, a director's note, and a timecoded fix list before you publish.
Four packaged installs built from the systems above — fixed scope, built on your stack, owned by you.
The AI sub dispatcher that lives in your phone: confirmations, rescheduling cascades, bid coverage with leveling, inspection corrections, and compliance paperwork — dispatched from your own number, with your approval on anything that matters.
korets.net/subpilotFinds the revenue leaking out of an HVAC, plumbing, or electrical shop — unsold estimates, missed calls, a silent customer base — and installs the system that recovers it: AI follow-up, 24/7 front desk, database reactivation.
korets.net/revenue-recoveryA dedicated operating layer connecting client communication, IRS transcripts, documents, tax software, and billing. It starts with one bounded workflow, keeps the professional tax suite in place, and expands only after the firm measures the economics.
korets.net/blackinkMost established companies won't survive a frontal AI transformation — the org resists it, and it should. The move that works: build the AI-first twin of one core process as a sidecar. Small, separate, no threat to the team that runs today's business — until the twin proves out and grows into the main line. I design and build the twin, install the human approval gates, and hand your people the keys. Fixed scope, measured against the process it mirrors.
Start the conversationIf it runs on phone calls, spreadsheets, and chasing people — it's a candidate for one of these patterns. Custom engagements start with a one-week audit of how the money actually flows.
Book a call