Frontier models are rented generalists. For the work your business repeats, a model trained on that work, and improved from its own traffic, is cheaper, faster and yours.
We research how to make that true in practice: how small a model can be and still hold a frontier model's quality on a task, how a model can keep learning from real traffic, and what it costs to run AI at scale.
How small a model trained for one task can be while matching a frontier model on it, and what it costs to run. We train open models on one task at a time and score them against hosted proprietary models on public benchmarks, under one scorer.
Evidence
QMSum: a 1.2B model trained on Ertas scored 35.41 ROUGE-1 against 32.43 for the best of five hosted proprietary models (paper, 18 Aug 2026).
How production requests and feedback become training data, and how a new version earns its place. This is the method behind the adaptive endpoint, where each new version passes a customer's evals before it takes over.
Evidence
Running in design-partner pilots. Nothing is published yet; results will appear on our research page.
Where AI spend goes, and what it costs to run agents at scale. We read public spend data to find the tasks that repeat at volume and cost the most per token, because those are the tasks worth training a model for.
Evidence
OpenRouter analysis (12 Oct 2026): classification, extraction and transformation were 20.8% of classified model spend over 30 days.
Each design-partner deployment leaves something reusable in the adaptive endpoint and the Ertas platform: an evaluation harness, a data-preparation step, a serving improvement or a training recipe. The research picks the problems, the deployments prove the method on real traffic, and the product absorbs what works.
We publish methods, measurements, code and models on public benchmarks. Every figure carries its date and what it was compared against, and we correct our own numbers in public when a later run overturns them.
CEO, AI Research
AI builder and startup operator. Data analytics and ML research at the University of Melbourne. Leads the lab's research.
CPO, Distributed Systems
Leads inference engineering. Background in product design and HCI research at the University of Melbourne.
CTO, AI Systems & GPU
Computer scientist with 6+ years across ML, computer vision and LLM systems. Leads model training.
Working with us
Book a callResearch collaborations, grants and press
hello@ertas.ai