The flexible escape · 2
€1,465 · Full cash refund
See how a photo changes the answer.
Maya
Money back. No entrance stairs.
Ready to check
Explore a recorded result.
Build it one step at a time.
01Find a model
→02Teach your job
→03Test its answers
→04Add pictures
The AI reads the holiday. Plain code checks each traveller’s wish list.
Keep the complete guide beside you ↗Start with a model that can classify text.
“Cash refunds are available.”
ModernBERTMoritz Laurer’s zero-shot classifier
MeetsViolatesCan’t tell
Ask Claude
Look at this open classifier. Explain what it already does, check whether it fits my task, and tell me what I would need to run it locally.
Measure it before you train it.
Ask AI to get it running.
Ask Claude
Set up this repo on my computer and launch the travel demo with its supplied model. Check my machine, download the required files, and add image understanding if supported. Open the app and help me check one holiday. Keep the model version; ask before any paid setup.
Tell it the job you want done.
MayaCash back · No stairs
Booking termsFull cash refund
Ask Claude
Build a travel checker that reads a holiday’s booking terms and checks four things: a full cash refund, arrival after midnight without calling ahead, a pool guests can use for free, and an included guided hike. Answer yes, no, or can’t tell for each. Show me examples to check before training. Check photos separately from the written terms.
Give Claude the complete brief.
1. Define the job before writing training code
My input is [a holiday offer's complete written booking terms]. My questions are [full cash refund before the deadline; check-in after midnight without arranging it; included pool access; included guided hike]. For each question choose exactly one of [yes, no, can’t tell].
2. Inspect my computer and the starting model
Inspect available memory, accelerator, disk space, Python and installed tools. Explain what can run locally and what is untested. Do not rent a GPU, buy a service, add paid calls, or download unrelated models.
3. Make a safe working folder and checked data
Preserve `apps/agency/models`, the frozen experiments, and the working persona app. Create a new workshop folder with `tools/tutorial.py prepare`. Keep every new output there. Never overwrite the supplied demo checkpoint or an earlier run.
4. Measure the model before changing it
Run the unmodified classifier on the development set and save predictions or a metrics receipt with the model and data hashes. Report accuracy and macro-F1 in plain language and show actual mistakes. Compare simple rules where they could solve this job. Do not assume fine-tuning is required or promise it will improve performance.
5. Train a small first experiment
Use `tools/tutorial.py` as the working travel recipe. Run an eight-decision smoke check in a separate folder before the full run. Say clearly that this checks the pipeline, not model quality. Then create a new folder for the full dataset.
6. Test it fairly and keep the failures
Freeze the selected checkpoint before the final test. Compare the original and trained models on identical held-out cases, using the same serialization and candidate descriptions. Include all errors, accuracy, macro-F1 and the individual predictions. Distinguish development progress from final-test performance. If the specialist does worse, say so and keep the report. Do not tune on final-test mistakes and call the same data a fresh exam afterward.
7. Run one new input, then connect it to an app
Show me how to load my saved checkpoint and classify a new JSON input. Map outputs to an inspectable result with an explicit review branch for missing or uncertain evidence. Start with a separate demo using the tutorial output. The existing travel API has fixed questions and calibration settings; do not point it at a new checkpoint without a compatible adapter and fresh validation.
8. Deliver a project I can run again
Give me one README with prerequisites, exact commands, expected files, troubleshooting and a first-result check. Include the base model pin, data provenance, code, checkpoint manifest, measurements, failures and upstream credits. Keep local inference, model downloads and the AI assistant's build-time costs separate. List what still needs my judgment.
Give it three choices.
“Refunds are hotel credit only.”
Which label fits?
Pair each example with an answer.
train-00000Open this example ↗
What it reads
“Contact the booking office to learn which cancellation rules apply.”
What we ask
Can I get a full cash refund?
Correct label
Can’t tell insufficient_evidence
Keep some examples for the final test.
Train it using checked examples.
Test the original classifier first. Use my checked examples to run a small training check, then train in a separate folder and save the best version. Keep final-test examples out of training, and show me the results and files.
Test it on examples it hasn’t seen.
Ask Claude
Compare the original model and my trained model on the same untouched examples. Show both sets of answers, the scores, and every mistake. Keep the report even if training made it worse.
Look at an answer it got wrong.
“arrival and registration at 01:30 are available automatically without any earlier contact or request”
Can you check in after midnight without arranging it?
Compare its answers with Jev's.
Same 360 fresh travel tests · 1,440 text decisions per model · V2 answers in about 0.1 s per holiday on this Mac
Our first model60.28%
Our V2 model95.28%
Jev98.61%
Jev scored higher. Our app now runs V2.
Ask it to try a new policy.
Cancellation policy
“Cancel seven days before arrival for a full refund to your payment card.”
New example
Ask Claude
Load the model we saved. Check whether this policy allows a cash refund: “Cancel seven days before arrival for a full refund to your payment card.” Show its answer, then try a policy that offers hotel credit instead.
Connect the models with code.
Aa
+Read the terms
Our trained text model
Look at the photos
Pretrained OpenJev vision
✓ ? ×
Check the wish list
The app combines the answers
Show how a picture becomes evidence.
Try a picture with the local model ↗Check what the photo can tell you.
Can guests swim here?
Is it a swimming pool?
Can guests use it?
Is it included?
This photo shows a pond. It doesn't show a swimming pool or tell us what guests can use. Check the booking terms for access and extra charges.
Check each person's requirements.
See where the costs come from.
Change one example yourself.
Your new dataset
No examples added yet.
You supply the answer. This editor does not train or predict.
Take the project and make it yours.
Your exampleTravel demo
Your starting pointModels + training code
Your instructionsPrompts + guide
The travel demo, prompts and guide are all in one repo.