AI that solves the problem, not just the demo.
I take a problem from the messy data all the way to something running in production — analysis, machine learning, LLMs and agents, evaluation, deployment, and the day it has to keep working without me. Peer-reviewed research behind part of it, and a systems-administration background that shows in how I build.
The machine does the preparation — the drafting, the searching, the analysis, the repetitive hours. The person keeps the decision, and what passes between the two is a reviewable artifact. That is what makes AI usable where the work is real: public administration, regulated data, teams with no ML department.
See how that works in practice →One person across the whole stack: data pipelines, models, LLM and agent systems, evaluation, deployment, and the monitoring that comes after. The problem does not get handed off four times before anyone actually solves it.
See the full stack →Correspondence and documents that eat staff hours. Knowledge nobody can find. Analysis that takes a week by hand. Training material that costs developer-months to produce. Bring me the problem and I will tell you honestly whether AI is the answer, and what it would take.
See the problems I have solved →Every project on this site is written up in full — the architecture, the numbers, and the caveats I would rather state myself than have you find. Peer-reviewed publications on large language models and computer vision. Systems administrator by trade, which is why I build for the day it has to run without me rather than for the day of the demo.
Read the full story →
About
Where I come from, what I have shipped, and the research behind it.
Read the full story →Projects
Multi-agent pipelines, DeFi security, XR, and LLM fine-tuning.
View the work →Videos
Technical walkthroughs on AI, data engineering and the tooling.
Watch →Stack
The full range — LLMs and RAG through vision, data and deployment.
See the stack →Contact
Open to AI/ML work, collaborations, or a good technical conversation.
Get in touch →Systems background, applied to AI.
Systems administration taught me how things break. AI is teaching me how to build things that don't.
I'm a data scientist and AI systems builder working across the full stack — data analysis, machine learning, deep learning, and large language models, from training and fine-tuning through inference, evaluation, and deployment.
I currently work as an IT officer and systems administrator, managing network infrastructure and security. That hands-on systems background feeds directly into how I build and deploy AI systems in practice, not just in notebooks.
Recent work spans multi-agent LLM pipelines, DeFi security analysis, and VR/AR development, with peer-reviewed publications on large language models and computer vision.
I build the system, then stay the supervisor.
The work below has a shape in common: split a hard problem into stages a machine can run on its own, make every handoff between those stages inspectable, and keep a human in the loop where judgment actually matters — not everywhere, and not nowhere.
All projects →Planning, generation, replay, evaluation — the pipeline does the work. The job is deciding what it should be doing and knowing when it is wrong.
T-Rax — Multi-Agent Unity Code Generation
A LangGraph-style multi-agent pipeline — orchestrator, planner, and executor agents — that plans and generates C# Unity scene code…
The problemBuilding one XR training scenario costs developer-weeks, so the organizations that need drills most end up with three of them instead of a library.
Gmail Assistant
An orchestrated n8n agent that drafts context-aware Gmail replies with Claude, negotiates and books meetings against live calendar availability, and…
The problemRoutine correspondence eats staff hours — triage, scheduling back-and-forth, and no reliable picture of what is still pending.
SmartReco
A four-stage DeFi security pipeline combining Slither static analysis with REVM-based dynamic transaction replay. Tested across 17 major protocols —…
The problemStatic analyzers hand teams long lists of maybe — findings nobody can act on, and the alert fatigue that hides the real one.
Technical walkthroughs, recorded end to end.
The whole stack, not one slice of it.
Six areas, and the tools I reach for in each. Where a project or paper on this site backs an area, it is linked — the rest is work in progress or not yet public.
LLMs & Fine-tuning
Training, fine-tuning and evaluating language models end to end — dataset construction and quality ablations, parameter-efficient adaptation, quantization for CPU and edge inference, and serving the result behind an API.
Fine-tuning ablations →Agents & Orchestration
Multi-agent systems where planning is separated from execution, each agent has one job, and the handoffs are inspectable. Tool use, structured output parsing, retries and idempotency for workflows that run unattended.
T-Rax and the Gmail Assistant →RAG & Knowledge
Retrieval pipelines from chunking and embedding through reranking and grounded generation, and the knowledge-graph side of the same problem — translating natural-language questions into structured queries over a graph.
WAILS 2025 paper →Computer Vision & XR
Monocular 3D object detection, dataset generation and benchmark construction on the research side; hand-tracked interaction and scene generation for headsets on the applied side.
RoadSense3D and VR Hand Interfaces →Data & Analytics
The layer underneath every model — ingestion and streaming, cleaning and feature work, statistical analysis with the assumptions stated, and visualization that carries an interpretation rather than just a picture.
Real-time pipeline walkthrough →Deployment & Infrastructure
Getting systems into production and keeping them there: containerized ML infrastructure, GPU servers and CUDA compatibility, reverse proxies and TLS, plus the network-security discipline that comes from running municipal IT.
SmartReco security pipeline →Let's talk.
Open to hearing about AI/ML work, collaborations, or just a good technical conversation. Four ways to reach me — pick whichever suits.
The most reliable way to reach me. I read everything and answer properly.
marcosau.dev@gmail.comFull professional history, roles and education, kept current.
/in/marcosauGitHub
Source for the projects on this site, plus smaller experiments.
@Marco-SauYouTube
Technical walkthroughs on AI, data engineering and the tooling around them.
@MarcoSau1981