Applied data science & AI engineering
I build machine learning and generative AI systems — forecasting, NLP, and LLM-based retrieval — for companies that need decisions grounded in data, not demos that don't survive contact with production.
01Services
Four defined offers. Each one is scoped to what you actually need — tell me what you're working with and I'll quote it.
Internal knowledge lives scattered across PDFs, procedures and contracts, and people either can't find the answer or trust an outdated version.
An assistant that answers directly from those documents, with the source cited every time.
Contracts, invoices and reports get read one at a time, by hand, and inconsistencies surface late, when they're expensive.
Extraction and cross-checking that catches the inconsistency before it costs you.
Generic FAQ bots answer from a script instead of what the company actually documents.
A support assistant grounded in your real documentation, not a canned script.
Demand, inventory and resource decisions get made on gut feel or last year's numbers, when the historical data to forecast confidently already exists.
A forecasting pipeline built on your own data, validated and ready to support planning decisions.
02Work
One system, documented end to end — including the parts that broke, alongside a track record that spans forecasting, NLP, and applied ML. Distill is not a service. It is the evidence behind the architecture I sell.
A deterministic multi-stage pipeline that runs daily on GitHub Actions, with LLM reasoning at two controlled points — matching and resume tailoring. Evolving toward multi-agent. It discovers roles across five job boards covering 76 companies, filters them through deterministic eligibility rules, and generates tailored resumes.
Durable state handling keeps runs recoverable after interruption — a property that came from a real failure, not from planning.
XGBoost-based warehouse occupancy forecasting for a large pharmaceutical and consumer health company (LATAM).
Sentiment analysis and time-series systems for a large automotive group's financial services arm (Canada).
Sentiment and friction analysis on a Brazilian federal platform, continuing research developed with UnB and the Ministry of Health.
03About
I'm a data scientist and AI/ML engineer based in Brazil. I build retrieval systems, agentic pipelines and LLM applications for companies that need them running in production, not demoing well in a meeting.
What defines the work is defensive architecture: single responsibility, validation at boundaries, graceful failure, instrumentation, and idempotency by default. These are not a checklist applied at the end. They decide how a system gets designed.
AIRA Labs is the structure through which I deliver that work. Small by intent — you talk to the person who writes the code.
04Contact
Forecasting demand, making sense of documents at scale, building an assistant grounded in your own knowledge — or something that doesn't fit neatly into any of those. The starting point is the same. Tell me what data you have and what decision it needs to support.
hello@aira-labs.com