Sofya
Sofya fine-tunes Llama models to cut doctors' paperwork in Brazil
Sofya, a Brazilian healthcare AI company, fine-tuned Meta's open-weight Llama models to structure patient records and assist physicians during consultations. The models run on dedicated Oracle Cloud infrastructure in Brazil. Sofya reports up to 30% less time on documentation and admin per consultation, and a 90% average CSAT.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
up to 30%
Reduction in time spent on documentation and administrative tasks per consultation, as reported by Sofya1
90%
Average customer satisfaction (CSAT) score, as reported by Sofya1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Doctors spend a large share of their day on electronic health record paperwork instead of patients. Sofya, a Brazilian healthcare AI company, wanted to give that time back to physicians and hospitals in Brazil and Latin America.
What they built
Sofya built a set of AI tools for providers. The tools structure data from consultations, identify named items in clinical text, and answer questions. The Llama case-study page describes a clinical assistant that guides doctors through taking a patient history, flags concerns and helps draft care plans. Llama is Meta's open-weight model family, meaning the model files can be downloaded and run on your own servers.
How they did it
Sofya used Llama 3.1 405B, the largest model, as a teacher. It generated high-quality synthetic training data, using a technique called knowledge distillation plus self-reflection prompts. Sofya then fine-tuned smaller Llama models of 70B, 8B and 3B on that data. Fine-tuning means training an existing model further on your own examples. The smaller models respond fast enough for real-time use. The models run on Oracle Cloud instances hosted in Brazil, which Sofya describes as an extra security measure. Sofya also plans a Llama 70B agent that uses tools and retrieval, which is not described as live.
Results
Sofya reports a reduction of up to 30% in time spent on documentation and administrative tasks per consultation, and an average CSAT of 90%. It says providers see better workflows and patient care, and that Llama brought cost savings and higher accuracy, but gives no figures for those.
Lessons for your project
Use a large model to train small ones. Sofya used its biggest Llama model to generate training data, then fine-tuned smaller models that answer fast enough for live use.
Keep patient data under your control. Open-weight models let Sofya run on its own cloud instances in Brazil, which it treats as an extra security measure.
Aim at a task clinicians already resent. Sofya targeted paperwork, which takes hours of each working day for many doctors.
Treat vendor-published results as claims. Ask for the baseline and method behind a time-saving figure before you build a business case on it.
Caveats
- Both sources are published by Meta, the maker of Llama. Meta states that all results are self-reported and not identifiably repeatable. No independent party confirms the figures.
- The Meta blog gives no sample size or method for the 30% and 90% figures. The 30% is an upper bound ("up to").
- The Llama case-study page words the time saving as 30% less time on paperwork during patient visits. This case uses the blog's wording, which carries the qualifier "up to".
- The blog says Sofya is on track for 1 million consultations per month. This is a target, not a result, and is not used as a metric.
- The deployed date is the blog's publication date. The sources do not say when the models went live.
- The Llama versions come from the case-study page. The blog names only the model sizes.
Sources
- How Sofya is taking clinical reasoning to the next level with Llama
AI at Meta · 2025-03-05 · Vendor case study · ai.meta.com
- Sofya | Llama case studies
Meta Llama · Undated · Vendor case study · llama.com
Summarized in our own words from the sources above. Spotted an error? See corrections.