Instituto PROA
Instituto PROA cuts interview-prep report time from 30 minutes to 5 with Llama 3.1
Instituto PROA, a Brazilian non-profit that prepares young people for jobs, uses Meta's open-weight Llama 3.1 on Oracle Cloud to write Portuguese reports about employers and openings. Meta reports that each report now takes about five minutes of staff time instead of 30.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Instituto PROA helps young people from low-income families in Brazil find work. Before each interview, a candidate needs facts about the employer and the role. Staff once built these briefings by hand, searching several websites and compiling a document. That took about 30 minutes per report.
What they built
PROA built an assistant that talks to students in Portuguese. A student types a company name and receives a structured PDF report to prepare for the interview. The text is written by Llama 3.1, an open-weight model from Meta, meaning its files are published and can be run by any company. The tools run on Oracle Cloud Infrastructure, where PROA hosts its other solutions.
How they did it
The system works in three steps. It sends the company name to a search service and collects web results. It stores them in a knowledge base, which acts as the lookup layer for a retrieval-augmented generation (RAG) setup, where a model answers from fetched documents. Llama 3.1 then writes the report from that material, and the system turns it into a PDF. PROA's executive director says it chose Llama 3.1 mainly for easy integration with its Oracle environment. PROA is exploring richer reports with Llama 3.2 vision models.
Results
The post says report preparation fell from an average of 30 minutes to five minutes. It also says the assistant drove 60x program growth, to 35,000 students enrolled per year, and that staff now spend more time on one-on-one support. These are Meta's account of PROA's results. No independent source confirms them.
Lessons for your project
Automate the research, keep the people for coaching. PROA freed staff from compiling reports so they could give one-on-one support to students.
Choose the model that fits your existing cloud. PROA picked Llama 3.1 mainly because it fit its Oracle setup, where all its tools already ran.
Ground the output in fresh search results. The assistant retrieves current web information about an employer before writing, rather than relying on what the model already knows.
Check what drives growth. Program growth has many causes, so measure the time saved per report separately.
Caveats
- The only source is a post published by Meta, which supplies the model and has a reason to present results well. Figures come from Instituto PROA via that post.
- The 60x growth and 35,000 students are program-wide figures. The post credits the assistant, but gives no baseline year or period and other factors may contribute.
- The post gives no go-live date. The deployed date shown is the post date. It reports no cost, team size or timeline.
- The post does not say how report quality was checked.
Sources
- How Llama and Oracle are helping Instituto PROA kickstart careers for students in Brazil
Meta AI Blog · 2025-08-27 · Vendor case study · ai.meta.com
Summarized in our own words from the sources above. Spotted an error? See corrections.