“A good model with good data outperforms a great model with bad data.”

Datamule Labs is creating efficient intelligence. This requires:

1. A rich training dataset. (The SEC Corpus)
2. Software to clean and structure data at scale. (doc2dict, xml2table, ...)
3. Compute. (we have received about $200k in compute from AWS, GCP, CF, Porter, Daytona, Render, Vercel,...)
4. Infrastructure to run software at low cost. (We use a combination of serverless, spot, and on demand compute via AWS, CF, GCP)
5. Cost minimization of LLMs served by inference providers. (Structured Output)
6. Latency minimization of LLMs served by inference providers. (Parallelized untagged XBRL extraction)
7. Self hosting ML models.
8. Novel caching insights.
9. Finetuning & Post training LLMs.
10. More compute.
11. Pretraining & Mid training LLMs.

We have achieved (1-6).

Datamule was founded by John Friedman as an open source project in July, 2024. Incorporated as a startup in May, 2025.