Matías Battaglia Romano
Senior Cloud Architect
AI strategy, grounded in engineering
I'm a Solutions Architect specialized in AI and Cloud architecture. Over the years I've helped dozens of organizations design and run AI workloads in the cloud, and trained hundreds of engineers and decision-makers along the way. Lately my work has shifted toward strategy and adoption: helping companies define what AI means for them, take their first concrete steps, modernize existing processes, and handle the change management that real adoption requires.
- Generative AI
- ML
- AWS
- Cloud Design
- Enterprise Solutioning
- AI Strategy
My Certifications
AWS Certified AI Practitioner
AWS (last renewed/obtained: Aug 26, 2025)
Expires Aug 26, 2028
Talks I've given
Beyond the PoC: how to bring AI agents to production with Amazon Bedrock AgentCore
AWS Generative AI Day Madrid 2025 · Madrid, Spain · Nov 3, 2025
Quantization and Optimization for image generation models
AWS Summit Madrid 2025 · Madrid, Spain · Jun 11, 2025
Generative AI Security: Best practices & Strategy
AWS Summit Madrid 2025 · Madrid, Spain · Jun 11, 2025
Design principles for Agentic AI
AI Meetup - The Information Lab · Madrid, Spain · May 22, 2025
Intelligent Agents: Building the Future of Automation
AWS Generative AI Day · Madrid, Spain · Mar 25, 2025
Round table: Innovating with AI
Cisco Partner Executive Exchange II · Madrid, Spain · Feb 6, 2025
Things I believe in
- Work at the intersection of technology and business.
- The best outcomes come from translating fluently between both worlds, not picking a side. Strategy without engineering depth is hollow; engineering without strategic context misses the point.
- Start from first principles.
- Frameworks and libraries change every quarter; the underlying ideas don't. Understanding why something works compounds over time; that's the bet behind my blog.
- Technology is an enabler, not the point.
- The win is what people can do because of the technology, not the technology itself. If a simpler tool gets the job done, use the simpler tool.
- Decisions need data behind them.
- Opinions are cheap. Architectures, vendor choices, AI rollouts; the ones that hold up are the ones where someone bothered to measure.
- The best decisions hear every voice in the room.
- The engineer in the back, the analyst who's been doing this for ten years, the customer who'll actually use the thing. The good answers rarely come from whoever speaks loudest.
- The most exciting part of AI is who it lets in.
- I have friends and family with no technical background shipping web apps, games, and tools they couldn't have built five years ago. That's the real story, not what you see on social media.
Writing
Long-form notes on language models, retrieval, and inference, collected at matiasbattaglia.com.
-
Transformers from First Principles · Part 1 — Tokenization
How language models split text into tokens, and why every vocabulary choice is a tradeoff between coverage and cost.
-
Transformers from First Principles · Part 2 — Embeddings
Turning tokens into vectors. What positional embeddings encode, and why dimensionality has hard tradeoffs.
-
Building a RAG from scratch
Chunks, embeddings, similarity search. The full retrieval pipeline walked through with 3D visualizations of every step.
-
Optimizing Inference for Image Generation
Running FLUX.1 Dev at one-fifth of its required VRAM through quantization, offloading, and memory tricks.
Tokenizer Playground
Try this simple demo on LLM tokenization! Language models don't read text, they read tokens, and each model tokenizes differently.
o200k_base for GPT-4o, cl100k_base for GPT-4 / 3.5 and as a reference for Claude. Anthropic does not publish a tokenizer library; for exact Claude counts, use the count_tokens API. Prices illustrative.
About DoiT
None of the work above happens in isolation. I do it as part of my role at the Forward Deployed Engineering team at DoiT, serving thousands of organizations. Below, a few of their voices.
What they say
Cloudflow's new RDS End of Life alerts have allowed us to be more proactive on keeping our database instances up-to-date. The new solution gives us internal visibility ahead of time so that we can prepare for upgrades, instead of having to upgrade under pressure while incurring extended support costs.
Jon Fairbanks, Site Reliability Engineering Manager
What I really like about DoiT's approach is that you're very hands-on and proactive. Satyam would ping me a few times a sprint, letting me know about the most current features, checking in on how things are going. When we are going through a peak time, that proactiveness makes a real difference. Satyam always comes through whenever we need support and helps us leverage the right experts to get us where we need to be.
Chiamaka Ibeme, Engineering Manager, Platform
DoiT was a true partner, not a vendor. They helped us understand the problem, refine the vision, and build something production-ready far faster than we could have on our own. Their expertise, responsiveness, and commitment made all the difference.
Dr. Anish Kapur, Founder & CEO, Promptly
DoiT's Customer Success and Forward Deployed Engineering teams work very closely with us. The regular sessions with our CSM keep us focused on the right priorities, and the FDEs provide the deep technical guidance we need to validate decisions and optimize our environment. That combination has been genuinely valuable for us.
Alexander Lundberg Santos, Platform Engineer at Extenda Retail
When we started working with DoiT, we deployed Flexsave to save time and reduce complexity. We still use it today. But what really stands out is the expert support. Having someone to collaborate with on deep cloud cost topics, someone who really understands the nuances, is incredibly valuable.
Jesper Terkelsen, CTO at Monta
DoiT's Cloud Accelerator turned our AI idea into a shipped product, saving at least three months of development and delivering reliable, explainable insights our customers trust.
Scott Desgrosseilliers, CEO and co-founder, Wicked Reports
Working with DoiT gave us the speed of a startup and the rigor of an enterprise R&D lab. We launched an AI feature in ninety days with zero engineering lift, unified our cloud visibility across multiple products and built a foundation to scale AI responsibly.
Tim Green, SVP of AI, DaySmart
Our AWS environment had grown rapidly, and complexity was outpacing control. DoiT helped us bring structure and clarity, aligning financial and technical decisions through a shared foundation.
Maurizio Trezzi, Director of IT, Vivaticket
Four levers. Real math. No made-up benchmarks.
Usage Optimization
Rightsizing, automation, kill what you don't need.
How
Cloud Intelligence spots it. PerfectScale fixes it.
Data Efficiency
Snowflake, Databricks, BigQuery - warehouses leak money at scale.
How
We stop the leak.
Cloud Waste Elimination
Idle resources, zombie infrastructure, no governance.
How
We clean house across GCP, AWS, and Azure.
Rate Optimization
Commitments, pricing discounts, procurement leverage.
How
Commitment Manager runs the numbers. Our advisors close the deal.
Other tools generate recommendations. We implement them.
The team
Meet my fellow Forward Deployed Engineers.
Google Cloud AI
- Google Cloud AI/ML5
AWS AI
- AWS AI/ML3
AI
- AI/ML2


Pub
- Pub/Sub2


Zero
- Zero/Low-Downtime Migrations1

Other
- AWS Core Services20
- AWS14
- AWS Security and Identity14




+8
- AWS DevOps13
- AWS Networking13




+7
- Google Cloud Core Services12
- Google Cloud Kubernetes11
- Google Cloud Networking11
- AWS Kubernetes9
- BigQuery9
- Google Cloud DevOps9
- AWS Databases8




+2