Manuel Porras

AI Systems Engineer — AI automation and private, on-premise AI that runs unattended

Berlin, Germany · Available across the EU and remote (CET/CEST) manuelporras.com · github.com/porrasbrand · linkedin.com/in/manuel-porras

Profile

I have worked in digital marketing for US businesses since 2006. Today I combine that with AI engineering: I automate the business processes of agencies and their clients, consolidate data from ad and project platforms into dashboards, and build AI systems that run unattended.

My focus is systems that run on owned infrastructure — local inference, multi-agent automation, and the evaluation harnesses that prove output actually improved. For EU organisations holding data they cannot send to a third party, that is usually the difference between doing something with it and doing nothing.

I build complete systems, not prototypes. Everything runs in production with process management, error handling and monitoring.

Professional Experience

Breakthrough3x (USA) — US marketing agency for home-services businesses, mostly HVAC

2020 – present · remote · freelance

Head of AI Automation & Process Optimisation · 2025 – present

I own the use of AI across the agency's and its clients' operations.

• Meeting intelligence: Zoom transcripts turn automatically into action items, summaries and meeting prep in Slack and ProofHub (repo zoom-action-items, public)
• Central dashboard across meetings, client status and monthly reports; client scorecards that consolidate ad-platform data into Google Sheets automatically
• LLM-driven triage of ProofHub project tasks, in daily use
• Lead integration Meta Lead Forms → ServiceTitan for an HVAC client: enquiries reach the client's system with no manual step
• Per-client AI account experts that answer questions from call transcripts and project data
• All automation runs on a self-healing fleet of 22 agents (see Selected Work)

SEO, PPC & Automation · 2020 – 2025

• SEO, Google Ads and local search (Google Business Profile) for US HVAC and home-services businesses
• Automated reporting and recurring account-management workflows

Partner, Digital Marketing & Web Development — Awesome Medical (Las Vegas, NV, USA)

2006 – present · remote · freelance, together with a US business partner

Account management, SEO and lead generation for US aesthetic-surgery clinics (incl. Chicago, Milwaukee, Atlanta, Phoenix, New York, Las Vegas, Oklahoma).

• Clients with combined revenue of over USD 1.5M per month (as of 2018)
• #1 on Google (organic and Ads) for the main keywords of several clinic websites
• Up to 750 enquiries per month for one clinic group through content and SEO; Facebook campaigns delivering 160–190 enquiries per month per location
• Website relaunches, PPC strategy (Google Ads, Facebook, remarketing), landing-page optimisation; bilingual English/Spanish content for Spanish-speaking patients
• Since 2026 also AI tooling for existing clients, e.g. an AI expert agent for one practice and new practice websites

Selected Work

Full write-ups with architecture diagrams at manuelporras.com/work

Private Call Analysis Pipeline

Local transcription and analysis of voice data that cannot leave the premises

• 266 of 266 recordings transcribed, zero failures; 20.7 hours of audio
• Zero network egress — faster-whisper large-v3 on owned GPU hardware, local LLM
• 8–14× realtime throughput; PII redaction built into the pipeline, not bolted on
• Deterministic rules kept separate from model judgement so every decision-grade figure traces to source records
• Overturned the client's own working hypothesis, and could prove it

Self-Healing Agent Fleet

20+ AI agents running unattended across two machines

• Queue-based dispatch, true liveness detection (session-exists checks hide crashed agents), stuck-job recovery, blast-radius isolation between trust domains
• 22 managed / 22 up; 35 weeks uptime on the always-on fleet
• Recovers from reboots and crashes without intervention
• Ported across two environments; every divergence documented

LLM Evaluation Harness

Proving generative output improved — at volume, without self-bias

• Blind judge (never sees the source), positive controls, calibrated axes
• Comprehension 3.3 → 7.8; judge self-agreement r = 0.78
• 56 cycles documented, including every dead end — the harness killed an architecture I had recommended the same day, and caught a "fix" that made output worse

Technical Skills

AI / ML — Local LLM inference (ollama, faster-whisper large-v3), Flux/ComfyUI, RAG and knowledge retrieval, LLM evaluation and benchmarking, structured extraction, multi-agent orchestration, Claude Code, prompt engineering. Hosted APIs where appropriate: OpenAI, Anthropic, Gemini.

Backend — Python, Node.js, TypeScript, PostgreSQL, SQLite, MySQL, REST APIs, real-time systems, pgvector

Infrastructure — Linux, Docker, GPU/CUDA, tmux, cron, pm2, SSH orchestration, Apache, nginx, Cloudflare, GoAccess

Data / Web — Playwright, Crawlee, crawling at scale, SEO analysis pipelines, competitive intelligence

Integrations — Slack, Zoom, ProofHub, ActiveCollab, Google Ads, Google Analytics, Google Search Console, Google Sheets, Meta Lead Ads, ServiceTitan, Zoho CRM, WordPress REST

Domain — Performance marketing (SEO/PPC) since 2006, process automation for agencies, GDPR-compliant AI architecture, data residency by design

Education

M.A. Technology and Innovation Management — FH Brandenburg and Brandenburg University of Technology Cottbus, Germany · 2007 – 2010

Diplom Systems Engineer — Universidad del Valle, Cali, Colombia · 1994 – 2002

International Baccalaureate — Deutsche Schule Cali, Colombia · 1991 – 1993

Languages

Work Authorisation

Colombian national. German Daueraufenthalt-EU (EU long-term residence) with unrestricted work permission. EU intra-mobility rights under Directive 2003/109/EC. German citizenship application pending.