I'm Carlos, an AI engineer who builds with the people who run the business. After 20+ years in technology, mostly for financial services, I know what regulated teams need from AI: answers that are correct, actions that are controlled, and decisions that can be checked.
Open to Forward Deployed AI Engineer and Applied AI roles in financial services.
I designed an AI credit-scoring system for a fintech company that advances cash to businesses. An AI model reads each company's balance sheet from PDF filings, and a risk model scores it using that data plus credit-bureau records. Assessment time fell from 6–11 hours to about 20 minutes.
Happy to walk through the details in a conversation.
I built an AI assistant that checks how ready a company is for an ISO 9001 audit and finds where it falls short of the standard. Every answer is tied to a cited source, so the team can verify it. It cut assessment preparation time by 60% and turned preliminary evidence reviews from days into minutes.
Happy to walk through the details in a conversation.
RAG · Hybrid search · Cited sources
A job-search assistant where one AI drafts, a second AI reviews without seeing the first one's reasoning, and plain code verifies every claim before anything reaches the user.
The same safeguards a regulated team would ask for: independent review, verified claims, and a record of every decision.
I led the delivery of an AI system that sorts incoming service requests by their nature and sends each one to the right team. It replaced older rule-based classifiers with a language-model system. To test it, I built synthetic test data and a second AI that judges answers on hard edge cases. The correct category is now the first choice 97% of the time, up from 35%. Resolution time fell from 24 hours to under 2 hours, and compliance with routing deadlines improved by 17 percentage points. It runs with automatic regression tests and response-time tracking.
Happy to walk through the details in a conversation.
NLP · MLOps · Automated testing
I led the adaptation of a two-way, real-time voice translation system to specialized vocabulary, and built an end-to-end proof of concept connected to the company's call-center and CRM tools. Translated audio comes back in under a second. Global support availability rose by 30%, and agents could be assigned by skill instead of by language or location.
Happy to walk through the details in a conversation.
Speech AI · LLMs · Cloud integration
An assistant that answers from real documentation and shows its sources. Tested on 240 questions.
One interface for two AI providers, with every trade-off recorded.
Knowledge lifecycle automation for a vault of 13,000+ notes.
Master's thesis: a machine-learning pipeline on 393,000 readings from a petrochemical plant, with average error under 0.25% on test data the model had not seen.
What is the system allowed to do, and who answers when it's wrong?
In my MLOps Docs Agent, a feature meant to improve search made results worse. I measured it, saw that, and left it off.
Every important trade-off is written down, so the team can see why things are built the way they are.
20+ years in technology, mostly for financial services, from developer and architect to program lead.