01
Describe the goal
Write naturally. Laplace understands prediction, forecasting, clustering, segmentation, causal questions, and more.
Opening the workspace…
A research workbench for machine learningMachine learning that
Machine learning that
trains itself, from your data.
Upload a dataset, describe your goal, and Laplace builds, validates, explains, and packages the right model for you.
Private by default. No user data taken for training or resale.
laplace · new project
alloy_strength.csv
412 rows · 9 columns · profiled
Your goal
Plan
- Regression task
- Grouped validation
- Baseline + XGBoost + LightGBM
- Trust checks enabled
Result
Trust: Medium/HighXGBoostselected · R² 0.88 (5-fold)
Report, notebook, and prediction tool ready.
Trusted by researchers at
How it works
From a sentence to a model, in three steps.
02
Laplace builds the workflow
Laplace profiles your data, checks for problems, selects model families, trains candidates, and compares against baselines.
03
Get a result you can trust
Receive a model, report, notebook, reusable workflow, and a deployment endpoint when the trained model is eligible.
Breadth without overwhelm
One workbench. Many kinds of machine learning.
Laplace routes each problem to the right family of methods, instead of forcing every dataset into one model.
Predict
Classification, regression, and quantile prediction.
Discover
Clustering, dimensionality reduction, and anomaly detection.
Forecast
Time series, demand, sensors, and experiments.
Explain
Statistical tests, causal inference, and survival analysis.
Optimize
Active learning, Bayesian optimization, and next experiments.
See
Image classification (ResNet/ViT), CLIP zero-shot, detection, and segmentation on the GPU worker.
Read
Text classification (BERT/RoBERTa), summarization & translation (T5), and document search.
Hear
Audio classification (Wav2Vec2/HuBERT), speech recognition (Whisper), and anomaly detection.
Science
Graph neural networks, molecular, bioinformatics, geospatial, and reinforcement learning.
Open the full map of what Laplace can build across tables, text, images, audio, science, optimization, and deployment-ready workflows.
The model router
Laplace chooses the right approach for the data you actually have.
A small CSV, a messy lab export, a folder of images, a COCO detection dataset, an audio archive, and a time-series file should not be treated the same. Laplace routes each problem to the right modeling family.
Your goal
- TablesXGBoost · LightGBM · random forest · SVM · logistic regression
- ImagesResNet · ViT · CLIP · YOLO · segmentation
- TextBERT · RoBERTa · T5 · Sentence-Transformers · extraction
- AudioWav2Vec2 · HuBERT · Whisper · mel-CNN
- ScienceGNNs · molecular · reinforcement learning · causal
Accuracy is not enough
Trust before accuracy.
High accuracy means nothing if the validation is wrong.
Laplace checks whether the model deserves your trust. It looks for leakage, weak baselines, small samples, bad splits, imbalance, overfitting, and out-of-distribution predictions.
A typical tool
97%accuracy
No questions asked.
Laplace
caution97%accuracy
Warning: likely leakage.
Recommendation: remove the post-outcome column and rerun grouped validation.
Leakage detection
Baseline comparison
Group / time-aware validation
Small-sample warnings
Class imbalance warnings
Out-of-distribution checks
Conformal uncertainty
Plain-English trust verdicts
Private by default
No data taken from your workspace.
Laplace keeps uploads, prompts, reports, predictions, models, and workflows inside your workspace. We do not use your data to train foundation models or sell it. Publishing is always explicit.
Private workspace
Your datasets are Private.
Uploaded files and generated artifacts stay tied to your project. They are not pulled into training sets, sold, or published unless you choose to share them.
Private by default
No model training on your data
No selling user data
Organization-level permissions
Signed links
Deletion controls
Your dataYour workspace
Model + report, for you
Foundation trainingnever used
Public sharingopt-in only
Workspace accesscontrolled
Reproducible by design
Every model comes with a methods-ready record.
Laplace produces reports you can understand, share, reproduce, and cite.
Model card
alloy_strength · v3
- Task
- Tabular regression
- Best model
- XGBoost
- Validation
- Grouped 5-fold
- R² (held-out)
- 0.88 ± 0.03
- Trust
- Medium/High
report.pdf
Methods-ready PDF
report.md
Markdown report
notebook.ipynb
Reproducible notebook
model_card.md
Model card
bundle.zip
Reproducibility bundle
predict.py
Prediction script
citation.bib
Citation export
/predict
Deployment endpoint
Simple by default. Configurable when needed.
Let Laplace choose, or open the hood.
Laplace can choose for you. If you know exactly what you want, open advanced controls for model family, split strategy, metrics, epochs, GPU type, thresholds, and runtime.
What would you like to predict, explain, or discover?
Advanced controls
Model familyauto
Validation splitauto
Metricauto
Epochsauto
GPU typeauto
Runtimeauto
Random seedauto
Every override is recorded in the run, so the result stays reproducible.
Built for real datasets
Who Laplace is for.
Researchers
Turn experimental data into reproducible models and reports.
Labs
Preserve workflows after students and postdocs leave.
R&D teams
Build predictive models from experiments, sensors, and inspection data.
Students
Build credible ML projects without fighting infrastructure.
Technical users
Customize models without losing reproducibility.
Finance
Model risk, forecasts, and portfolios with auditable workflows.
Private beta
Free for research labs while we build it with you
There are no paid plans yet. Every beta lab runs on a Compute Passport — a monthly compute budget — and we fund heavy runs on request. Pricing arrives after the beta; beta labs help shape it.
Invite-onlyOnboarding labs in cohorts
What your lab gets
- Every model family — tabular, vision, text, audio, time-series, graphs & molecules
- Trust scoring, reproducible reports, model cards & a deployment endpoint
- $35/mo of compute, free — heavy runs funded on request
- The API and live prediction endpoints, included
Start with a dataset. Leave with a model you can explain.
Private by default. Reproducible by design.