I build the infrastructure your AI runs on.
GPU and generative-AI pipelines, model deployment, and LLM platforms that keep private data private. I work directly with your team, remotely, in English or Portuguese.
What I do
GPU and generative-AI pipelines
Job queues and schedulers that keep every GPU busy, inference stacks tuned for speed, and the services that turn a model into a product. On consumer cards or cloud instances.
Model deployment (MLOps)
A repeatable path from a trained model to production: CI/CD that validates each build, containerised inference on cloud GPUs, and tools to review what the model actually does.
LLM platforms for private data
LLM gateways that mask personal and health data before a prompt leaves your systems, so AI features stay within LGPD and HIPAA limits.
Case study: a generative AI platform for game assets
From 2024 to 2025 I was Principal Engineer at Sortium, which generated game assets with AI: diffusion-based PBR textures up to full 3D assets, on Azure and an on-premises GPU cluster. The product launched; the company closed in 2025.
I defined the architecture end to end, from the generation pipelines to the customer-facing services. I designed the queue service and scheduler for a cluster of 16 RTX 4090 cards, balancing 5-second texture jobs against 90-second 3D jobs, and packed every model into each card's 24 GB so any GPU could take any job.
The same kind of pipeline now produces the asset packs of Lab278 Studio.
| Measure | Before | After |
|---|---|---|
| Single-material texture | 6-10 s | 3-6 s |
| Multi-material asset | 50-60 s | under 10 s |
| Worst-case capacity | 5,760 jobs/h | 9,600 jobs/h (+67%) |
How an engagement works
A first call
We talk through your system, the problem and the constraints.
A written proposal
Scope, approach, timeline and price, in a short document you can share inside your company.
The work
I work in your repositories and channels, in small reviewed steps, and you see progress every week.
Handover
Documentation and a walkthrough, so your team can run and change what we built.
About me
I'm Carlos Tosta, an engineer in Rio de Janeiro with more than 15 years of production systems, from national government platforms to AI startups.
For 13 years I worked at Dataprev, Brazil's federal IT company, where I led the research and development of bCPF, a permissioned blockchain for the national CPF registry launched with Receita Federal, and later managed an architecture division.
After Dataprev I was Principal Engineer at Sortium, built the model deployment path for Phood's kitchen vision system (through Buoy), and worked as founding engineer at HealthForge, a HIPAA-compliant telehealth platform.
Have an AI system to put into production?
Tell me what you are building, and we will see together whether I can help.
Or write to consulting@lab278.com