Mario Alexandre

I build the infrastructure that makes large language models work in production: continuously updating model substrates, multi-body adapter architectures with empirical retrieval gates, and the tooling that holds it together. I came to AI from electrical engineering, and I treat a model the way I was trained to treat a signal or a control loop, as a system you measure, bound, and prove, not one you hope about.

I founded sincllm.com and built the AI engineering practice and the full technical stack under it. In parallel I build and run the AI systems that operate companies generating millions in revenue, one of them serving 5,000+ enrolled users, 300K+ subscribers, and 155+ properties every day.

Why an electrical engineer builds AI systems

Find the signal hiding in the noise. That instinct does not care whether the system is a radio channel or a language model.

What pulled me into engineering, and then into AI, is one feeling: walking into a field of pure noise and finding the signal buried in it. That is what logic actually is to me. Not a flash of genius, but the patient refinement of one step after another until the path to the goal resolves out of the static. Signal processing, control theory, and sampling taught me to do that with circuits and waveforms. A language model turned out to be the same problem at a larger scale: a vast, noisy space where the real work is shaping the input, bounding the behavior, and refining until the right answer separates from everything else. A prompt is a signal you can undersample. A model that drifts is a control system without a feedback loop.

I came to this from electrical engineering. I hold a BS in Electrical and Electronics Engineering from the University of South Florida, my senior capstone designed a bidirectional DC/DC converter with MOSFET control and researched OFDM hybrid waveforms for 5G, and I keep a curated knowledge graph of the EE sources I still reason from. Then, in my last semester, I took a class on artificial intelligence, and that was the moment it clicked. I fell in love. Everything I already loved, the sampling, the feedback loops, the discipline of proving a system before trusting it, was right there again, only now it was pointed at a machine that could reason.

The honest answer to the computer science versus electrical engineering question is that the work below needed both. The sampling theory that makes OFDM work is the same math under an adapter architecture, and the specification discipline that makes a system pass review is the same discipline taught in EE senior design, not in most ML bootcamps. The systems instinct never changed. Only the domain did.

What I have built

Every project below shipped with measured results, traced from specification to deployment under Ousterhout interface-first discipline. I have authored 30+ Software Requirements Documents this way.

• Ranked a client #1 on Google. Built a custom SEO engine that took a client’s pages to the number one organic result on Google, then automated the technical SEO (indexing, schema, internal linking) so the ranking holds without hand-tuning.
• 200+ articles a month, fully automated. Engineered transcript-blaster-v2 (Python 3.12, FastAPI, 67 passing tests, deployed via systemd to a VPS), the pipeline that researches, writes, and distributes more than 200 on-brand articles a month for a single company, hands-off.
• 50,000+ search-engine bots a month. Grew a site’s crawl to more than 50,000 Google and AI-crawler visits a month through schema, sitemap, and IndexNow engineering, the prerequisite for getting indexed, cited, and found.
• An AI SMS engine that sends 1,200 texts an hour. Built and deployed a production outreach system with AI-generated, per-recipient messaging that runs at 1,200 messages an hour on a $50 a month plan.
• AI that operates real web apps. Built a phantom-style automation framework that drives live browser sessions through genuine clicks and typing, human-like and session-persistent, so it runs unattended where brittle scripts get blocked.
• RAG that grounds ChatGPT and Claude. Built an advanced retrieval system (a 75-book corpus, 555 chunks indexed, BM25 and dense fusion served on vLLM) plus a machine-readable llms.txt layer, so external models fetch a site’s own knowledge and answer accurately and on-brand instead of guessing.
• A private, verified source of truth. Built a local-model RAG that extracts a company’s data, distills it into procedures, and trains a self-hosted model on them, with a QA agent checking every step. The result is one private, accurate source of truth that never leaves the company’s infrastructure.
• 70 to 80% lower AI bills, quality held. Built a multi-layer system that runs most work on local models on the company’s own hardware and routes only the hard calls to Claude, Codex, or the cheapest capable cloud model. The AI bill drops 70 to 80%, the data stays private, and the quality stays maxed.
• The research under all of it. The harder infrastructure that makes the above reliable: continuously updating model substrates, multi-body adapter architectures with empirical retrieval gates, and rank-one model editing measured at a 92% localized-edit hit-rate across 75 paired tasks and 10 of 10 cross-domain composition, every system specified and validated before it shipped.

Education

The through-line across every project above is the same instinct an electrical engineer develops by necessity: measure the system, bound its behavior, prove it works before shipping it. That is what I bring to LLM infrastructure.