Luka Meng
ships the AI systems revenue teams can actually run on

B2B Partner Portal — BetterBusiness

The system that decides which partners receive which leads, in which postcodes, under which filters — and bills for them. Inherited as a vibe-coded production app with no version control, no tests on the money path, and no deployment pipeline. Migrated to GitHub/Vercel, replaced a brittle Zapier ingestion path with canonical edge functions, and rebuilt the delivery write path around transactional, fail-closed operations behind an enforced migration release gate. Most implementation is done by a fleet of AI coding agents; my work is the judgment layer — scoping, adversarial review, release gating, and verification against production.

Agent Fleet Governance & Skill Infrastructure

The operating layer that makes a fleet of AI coding agents safe to point at production. Reusable skills are authored once and installed into two different runtimes through thin adapters, so behaviour can't silently diverge; a provenance checker verifies the mirrors in CI. Consequential work goes through cross-provider adversarial review — Codex's reviewer is Claude, Claude's reviewer is Codex, and each provider's reviewer is deliberately not installed in its own runtime, because self-review isn't review. Every writable lane gets its own isolated worktree, and a separate judge assesses each planned action for irreversible risk and whether the transcript actually authorized it.

Production Voice-Agent Release Engineering

Release discipline for AI voice agents that qualify real customers in German, where the failure mode isn't a crash — it's an agent that proceeds without properly obtained consent. Built a default-deny test control plane where a call is only judged if it can be exactly attributed by provider ID, so an unrelated production conversation can never be mistaken for test evidence. Scenarios are named and single-use; the harness stays disarmed outside an approved call. Post-publication monitoring reads real calls on both sides of the risk — false negatives and newly introduced false positives.

Multi-Agent Orchestration Platform

Three AI agents coordinating autonomous tasks across platforms without human relay — signal ingestion, enrichment, and 5-stage triage. Built on a self-hosted OpenClaw gateway (Hetzner VPS) routing Kimi K2.5 with OpenAI fallback. Evolved through real production failures: zombie cron race conditions, context window overflow fixes, halt-on-failure dedup. Every safeguard exists because something broke. Retired April 2026 after the BetterBusiness offer closed its purpose.

AI Lead Qualification Voice Agent

AI voice agent that qualifies inbound leads via automated phone calls, replacing a 3-touch manual outreach cycle (10–30 min/lead). Handles interest confirmation, language detection with human handoff for French-preference leads, deal qualification, and booking link delivery via WhatsApp and email.

Speed to Lead Automation

End-to-end Make automation replacing a 5–10 min/lead manual outreach workflow. Pulls Salesforce leads, matches contact and program data, generates a personalized first-touch email, sends via Outlook, and logs back to CRM. Formal IT proposal submitted for Salesforce API and MS 365 OAuth access to enable company-wide deployment.

Sales Process Redesign: Interview & Campus Visit Pipeline

Restructured interview and campus visit scheduling from a disconnected, email-heavy process into a pipeline stage with full visibility. Self-service booking via Microsoft Bookings eliminated ~10 emails per lead; Power Automate dashboards gave the team real-time handoff visibility. Standardized feedback system with auto-generated PDF reports made interviewing delegatable to regional teams for the first time.

Extended the same infrastructure to campus visits — the in-person equivalent of a product demo, where prospects tour the kitchens, meet faculty, and experience the environment. For programs up to €67k, this is where the buying decision happens. Automated reminders near-eliminated no-shows.

AI-Enabled Enrollment Document Processor

Enrollment applications arrive as multi-page PDF bundles (transcripts, passports, motivation letters, financials) that sales reps must manually split and classify before acting — ~15 min per file. During peak intake periods, files sit untouched while reps are tied up with other leads. Built a Power Automate pipeline: AI Hub classifies document types and extracts fields, Adobe PDF Services API slices pages, and each document routes automatically to the right handler.