A continual-learning lab
AI that gets better at its job by doing it.
We build models and harnesses that learn new tasks, keep what they already know, and improve inside the enterprises that use them.
The problem
Today, an AI system is trained once and then frozen.
A company should be able to teach its AI a new task, correct its mistakes, and watch it build an understanding of the business that compounds. That almost never happens. Models are fine-tuned once and left alone. The software around them is written by hand and rarely changes. Every correction an expert makes is gone the moment the session ends.
Capability is also getting cheap. When everyone can rent the same frontier model, the advantage moves to what only you can learn — and those lessons live in your own work, not on the open web.
Our approach
Two halves of one system, learning together.
01 · The model
Learns new capabilities
Not a one-off fine-tune, but a repeatable process of learning, evaluation and controlled updates — new tasks learned while earlier skills, and the ability to keep learning, are retained.
02 · The harness
Learns how to do the work
The tools, memory, checks and workflows around the model learn from successful outcomes and expert corrections, propose better ways of working, and test those changes before they are used.
A better harness produces more useful experience to learn from. A better model takes on harder work and improves its own workflows. That recursive loop is what we are building.
What we have shown
Early, measured, and published with its caveats.
#2JevBench · Jev-class
Hopper, our 4B decision model, ranks second in JevBench's Jev-class capability ranking — 63.5 against the leader's 64.7 — with the highest calibration of any system in the top six. It was built within a week of the benchmark's release.1
Top 5Decision Index · under 5B
On the Jev Decision Index, a second independent leaderboard that scores 70 open systems across 43 benchmarks, Hopper ranks third of the 46 under five billion parameters.2
+14.1%Tool use · internal
Our update method lifted a tool-use evaluation from 57.9% to 66.1% without losing earlier abilities on the retention suites we track.3
2Public benchmark suites
Before this, we built open models and public benchmark suites for fashion retrieval and attribute extraction — with frozen protocols, predictions hashed before labels open, and the runs we lost published beside the ones we won.4
1JevBench v1.4.2, 24 September 2026 snapshot, scored by an independent maintainer. Second on capability; fifth on the composite score, which also weighs speed and cost. Build time is team-reported. The ranked model is released for research and demonstration.
2Jev Decision Index 0.2.1, updated 28 September 2026 and maintained independently on Hugging Face. Hopper (G) 1.2, 4.66B served parameters. Third of 46 on the chance-corrected Decision Index at 40.77, within 0.1 of fourth; 18th of 70 overall. The edition is 76% scored.
3Internal result: 68 questions, mean of 16 samples, one seed. We will publish the protocol and artifacts before treating it as established.
4Protocols, scorers and prediction files at hopit-ai.github.io.
Customers
Early customers, and the conversations that come next.
Working with
- A D2C roll-upCommercial and trend intelligence API
- A fashion-tech startupStyling API
- A health-tech startupAn in-house System One decision model
In active discussion
- A Fortune 500 companyContinual learning
- A PE-owned unicornContinual learning
We do not name customers publicly.
Work with us
Most enterprises don't need a bigger model. They need one that learns their work.
We deploy with you. A small forward-deployed team works inside your environment, trains models for your workflows, and builds the harness that makes them useful and keeps them improving.
Your data and your deployed models stay yours; we bring the methods and the infrastructure behind them. That makes the approach ideal for regulated enterprises — banks, lenders, insurers and healthcare providers — where data cannot leave the premises and people stay responsible for consequential decisions.
- Choose one workflow and agree on what success means.
- Prove it inside your environment, against that agreed measure.
- Expand from the first deployment to the workflows around it.
Research log
We publish as we go.
-
Hopper places top five under 5B on the Jev Decision Index
A second independent leaderboard: 70 open systems across 43 benchmarks. Hopper (G) 1.2 ranks third of the 46 systems under five billion parameters on the headline score.
-
Hopper places #2 in JevBench's Jev-class ranking
A 4B decision model built within a week of the benchmark's release, with the highest calibration in the top six. Research weights are open.
-
Continual learning with retention: +14.1% on tool use
The first internal run of our update method. Retention held across the suites we track. Protocol and artifacts to follow.
-
MODA_NER: an open benchmark for fashion attribute extraction
Four frozen tracks, never averaged. Predictions hashed before labels open, and every published number recomputable from the shipped artifacts.
-
MODA: open models for fashion retrieval
Full-corpus evaluation through one harness, competitors included, and the benchmarks we lose shown alongside the ones we win.
Join us
A small team building AI that keeps learning.
We are researchers and engineers, and we still train models ourselves. We are looking for people who would rather measure a result than announce it — to push continual learning forward, and to take it into the enterprises that need it.
Write to us