Key Takeaways

  • General frontier models fail silently on recall when parsing complex documents in high-stakes fields like finance and healthcare.
  • Reducto cut its structured extraction cost by more than 50%, dropping from over 4.5 cents per page to under 2.25 cents with its custom R1 model.
  • Post-trained, task-specialized models beat generalist foundation models across all three dimensions of the Pareto frontier: accuracy, cost, and latency.
  • Merging layout analysis and last-mile error correction into one custom architecture beats chaining separate computer vision scripts and LLM calls.
  • Renting foundation models works for broad horizontal features, but owning proprietary models becomes mandatory as volume approaches one billion pages monthly.

The Failure of Commercial APIs at Scale

When founders start building AI products, renting intelligence via commercial APIs seems like the only rational choice. You write a few prompts, set up an API key, and ship. Reducto followed a version of this playbook in its early days. As Adit explains, “We started with this idea of okay, we'll use a lot of smaller traditional CV models for the baseline and then we'll use larger models for the really hard tasks.” (play from 4:45)

That hybrid approach broke down as production volume climbed. General frontier models struggled with document recall, silently missing lines of tables or fine print on financial statements. In industries like healthcare and lending, a silent omission is a critical bug. To make matters worse, chaining traditional computer vision tools to general-purpose LLMs created compounding latency and bloated expenses. Running structured extraction through off-the-shelf commercial APIs cost more than 4.5 cents per page. When you project that spend across hundreds of millions of pages, the unit economics collapse.

The Triple Win of Specialized In-House Architecture

The standard narrative in Silicon Valley says foundation models improve so fast that building proprietary models is a fool's errand. Adit's data contradicts this assumption. When Reducto engineered its own R1 model, it collapsed layout detection and error correction into a single pipeline instead of chaining multiple distinct steps.

“We keep finding that if you are looking at a specific type of task, a post-trained model on that task outperforms on all three dimensions,” (play from 6:19) Adit noted. R1 proved that founders do not need to trade speed or precision for margin. “R1 is just you send us the documents, it will be less than the per page cost of our base models for better accuracy than our most accurate models.” (play from 10:30)

By focusing the model architecture entirely on document parsing, Reducto slashed per-page extraction costs by half while eliminating the extraction errors common in general models. General models still hold value for wide-open reasoning tasks, but their role shrinks inside specialized operational pipelines. “Our stack has shifted more and more towards in-house over time,” (play from 6:30) Adit observed. “There are still really horizontal tasks where a foundation model makes sense, but that's limited in scope today.” (play from 6:34)

What to Do With This

Audit your production API spend this week. Pull the logs for your highest-volume prompt and run 200 production samples through an open-weight, domain-tuned model in an isolated test environment. If your API calls cost more than three cents per unit and you process over 100,000 units monthly, benchmark fine-tuning an open-source model against your existing commercial API baseline. Track recall and latency side by side to see if renting intelligence is silently eroding your gross margins.