Langsat Data Science Predict anything from any data

BUILT FOR ML ENGINEERS

BUILT ON ANALYTICS

Every Data Analytics feature, plus the model layer.

Everything you can do in Langsat Data Analytics — chart catalog, auto-gen dashboards, drill-down, themes, conditional formatting, edit prompt + explain, PNG / CSV export — is available here too. Click any tile to jump to the matching section.

Chart catalog Auto-gen dashboards Drill-down Themes Conditional formatting Edit prompt + Explain PNG / CSV export

INPUT SHAPE

One table or many — Langsat takes either.

Drop a CSV and you get a tabular model. Drop two or more linked tables and Langsat detects the foreign keys, builds a heterogeneous graph, and trains a GNN on the joined structure. Primary keys, foreign keys, and timestamp columns are auto-detected — and you can override any of them.

Tabular One table · no joins

A single CSV or table.

Langsat detects the primary key and timestamp automatically, then trains a tabular model — MLP, FT-Transformer, ResNet, or TabNet — on the rows directly.

Auto-detected: primary key (singleton, enforced), timestamp column (≥80% parseable datetime values triggers it).

Relational N tables · FK edges · GNN

Two or more tables, linked.

Upload your customers, orders, and products — Langsat detects foreign-key relationships, builds a heterogeneous graph, and trains a GNN (SAGE / GATv2 / GIN) directly on the joined structure. No flattening.

Auto-detected: foreign-key edges (dtype-matched to target PK), per-table timestamps, table relationships.

HOW IT PREDICTS

Four modes, picked for the question you're asking.

Every chat message and API call carries a mode. Pick one explicitly — or leave it on auto and Langsat chooses the right strategy from the question.

auto ●

Let Langsat pick the best approach.

Default for Data Science projects. Langsat reads the question, picks transductive when you're asking about a known entity and inductive when you're asking about a new one.

explore ●

Pure data analysis — no model in the loop.

Aggregations, distributions, and chart questions that don't need predictions. This is the only mode available on Data Analytics projects.

transductive ●

Score an entity that was in the training set.

Look up an existing customer / product / node by ID and ask the trained model to re-score it with the latest features. Fast lookups.

inductive ●

Predict for entities never seen before.

Pass new feature values, optionally with related rows or references to existing graph nodes — the GNN encodes the new entity on the fly. Two relational variants: inline data and existing-graph links.

Data Analytics projects lock to explore; Data Science projects unlock all four.

THE TRAINER

Ten models in the registry. Pick one — or race all of them.

Six tabular architectures (MLPs, FT-Transformer, ResNet, TabNet) and four relational ones (SAGE, SAGE-full, GATv2, GIN). Free tier ships the fastest model automatically; Pro tier lets you pick; Max and Team unlock benchmark mode.

Tabular · 6 models Relational · 4 models

MLP Small (64ch)

64 hidden channels — fastest, cheapest.

MLP Baseline (128ch) ★

128 channels — Pro tier default. Solid baseline.

MLP Large (256ch)

More capacity for complex tabular targets.

FT-Transformer

Attention over your feature columns — strong on heterogeneous tables.

ResNet

Residual blocks for deeper representation without overfitting.

TabNet

Attentive feature selection — picks what matters per row.

GraphSAGE ★

Simple neighborhood aggregation — Pro tier default for relational data.

GraphSAGE+

Same base, with extra regularization for noisier or sparser graphs.

GATv2

Graph attention — learns which neighbors matter most for each node.

GIN

Graph isomorphism — strongest at detecting structural patterns.

Free tier ships the fastest model in each mode automatically (no choice). Pro tier lets you pick any one model per training run.

Benchmark mode · Max / Team

Don't pick a model — race them all.

Flip the switch and Langsat trains every model in the registry in parallel, then saves the single best plus an ensemble of the top three finalists. Uses more credits — but no more arguing about which architecture is right for your data.

Benchmark ON

Training all models — best + ensemble of top-3 saved (uses more credits)

WHAT YOU CAN PREDICT

Supervised or unsupervised — six task families, every common subtask.

Three supervised tasks (regression, binary classification, multi-class classification) and three unsupervised (clustering, anomaly detection, link prediction). Tell Langsat the target column and it figures out the rest.

SUPERVISED

Regression

Predict a continuous value on a real-number scale — revenue, latency, lifetime value.

SUPERVISED

Binary Classification

Predict a yes / no outcome — churn vs retain, fraud vs legit, click vs skip.

SUPERVISED

Multi-class Classification

Sort entities into three or more categories — product type, segment, risk tier.

UNSUPERVISED

Clustering

Group entities by similarity — no labels needed.

UNSUPERVISED

Anomaly Detection

Flag the rows that don't look like the rest — fraud, outage, defect, outlier.

RELATIONAL

Link Prediction

Predict edges that don't yet exist in the graph — recommendations, friend suggestions, supplier matches.

MEASURE WHAT MATTERS

Every task gets the metrics that actually fit it.

A regression model wouldn't tell you a useful AUC; Langsat doesn't show one. Each task ships the metrics it earns — on a dashcard you can pin, theme, or export like any other widget.

Model overview · Regression

mlp_baseline · 18 features · trained 4 min

R² 0.872 MAE 3.42 RMSE 5.18 R² (log) 0.81

Training & Validation Curves

train val

Model overview · Binary Classification

ft_transformer · 24 features · churn target

AUROC 0.937 Accuracy 0.891 F1 0.864 Precision 0.882 Recall 0.847

ROC curve

Confusion matrix

Model overview · Multi-class Classification

tabnet · 16 features · 5 classes

Accuracy 0.823 F1 macro 0.798 Precision macro 0.812 Recall macro 0.785

Confusion matrix · 5 × 5

Model overview · Clustering

kmeans · 12 features · 4 clusters

Silhouette 0.642 Davies-Bouldin 0.81 n_clusters 4

Cluster projection · PCA 2D

Model overview · Anomaly Detection

isolation_forest · 14 features · 1,447 anomalies

contamination 2.4% threshold -0.082 n_anomalies 1,447 n_normal 58,820

Score distribution · anomalies highlighted

Model overview · Link Prediction

sage_baseline · users—rates—movies edge

AUROC 0.892 val_auc 0.881 edges 2.4M

val_auc · per epoch

Predicted edges

FROM MODEL TO PRODUCTION

One endpoint. Any mode.

Every trained model gets a REST endpoint instantly. POST a single prediction or batch up to 1,000. Same auth header across modes. Same response shape. Inductive when you have features, transductive when you have an entity ID.

cURL Python

POST /api/v1/predict

# single prediction · inductive mode
curl -X POST https://api.langsat.ai/api/v1/predict \
  -H "Authorization: Bearer rdl_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model_id": "8f3c-...uuid",
    "mode": "inductive",
    "features": {
      "country": "SG",
      "tier": "Pro",
      "signup_age_days": 412
    }
  }'

# entity lookup · transductive mode
curl -X POST https://api.langsat.ai/api/v1/predict \
  -H "Authorization: Bearer rdl_YOUR_API_KEY" \
  -d '{"model_id":"...","mode":"transductive","entity_id":42}'

# batch · up to 1000 items
curl -X POST https://api.langsat.ai/api/v1/predict/batch \
  -H "Authorization: Bearer rdl_YOUR_API_KEY" \
  -d '{"model_id":"...","mode":"inductive","items":[...]}'

# single prediction · inductive mode
import requests

resp = requests.post(
    "https://api.langsat.ai/api/v1/predict",
    headers={
        "Authorization": "Bearer rdl_YOUR_API_KEY",
    },
    json={
        "model_id": "8f3c-...uuid",
        "mode": "inductive",
        "features": {
            "country": "SG",
            "tier": "Pro",
            "signup_age_days": 412,
        },
    },
)
result = resp.json()
print(result["prediction"], result["probability"])

Endpoints

POST /api/v1/predict
POST /api/v1/predict/batch · max 1000
GET /api/v1/predict/models

Auth

Authorization: Bearer rdl_…

Per-project key. Rotate from the API panel.

Response shape

{
  "prediction":      <value>,
  "probability":     0.87,        // binary
  "probabilities":   { ... },     // multiclass
  "cluster":         3,           // unsupervised
  "anomaly_score":   -0.092,      // anomaly
  "is_anomaly":      true,
  "warnings":        [],
  "monitoring":      { ... }
}

INPUT MONITORING

Drift, nulls, unseen IDs — caught before they hurt.

Langsat tracks every feature you send through /predict. Distribution drift (KS / Jensen-Shannon), null-rate creep, unseen ID rate on key columns. Per-feature thresholds, three severities, in-app alerts plus integrations.

Input Monitoring

Last 7 days ▾

API Monitoring AI Assistant

churn_v3 · binary_classification · Last 7 days · 22,184 predictions

8 stable 2 warning 1 critical

Feature	Type	Health	Requests	Issues	Drift	Null %
MOVIES · 5 features
duration_min	numeric	ok	8,421	0	0.018	0.2%
language	categorical	warning	8,421	1	0.083	1.4%
title	text	ok	8,421	0	0.005	0.0%
release_year	numeric	ok	8,421	0	0.022	0.0%
rating_avg	numeric	ok	8,421	0	0.014	3.1%
RATINGS · 1 feature
rating	numeric	critical	22,184	2	0.221	0.0%
VIEWERS · 4 features
age_group	categorical	ok	22,184	0	0.028	0.0%
country	categorical	warning	22,184	1	0.068	0.0%
subscribed	boolean	ok	22,184	0	0.011	0.0%
session_count	numeric	ok	22,184	0	0.034	0.8%

Critical RATINGS · rating drift score 0.221 exceeds threshold (0.150). Likely cause: new tier rolled out in week 28.

Recommendation: fine-tune from current checkpoint with the last 14 days of data (delta 4.2% of parent — safely incremental).

FINETUNE

Fine-tune inherits the parent's entity table, FK relationships, and cleaning rules — you just point it at new data. It starts from the last checkpoint and trains on the row delta. If the delta exceeds 10% of the parent, Langsat warns you to consider a fresh retrain instead.

Step 1 · Upload

Drop in the new data.

Langsat auto-routes each file to its matching training table.

＋ Drop CSV / Parquet here

orders_oct.csv 2.4 MB · 8,421 rows

reviews_oct.csv 1.1 MB · 4,892 rows

Step 2 · Verify

Schema check + delta count.

Same cleaning rules as v1. New rows only — no overlap.

Schema matches v1 · 18 features

Delta

13,313 new rows

4.2% of parent — incremental OK

Δ > 10% would warn "consider fresh retrain"

Step 3 · Fine-tune

Resume from the last checkpoint.

Inherits pipeline config from v1. Trains only on the delta.

parentv1 · sage_baseline

taskbinary_classification

max duration30 min

cost~42 credits

Fine-tune lineage 8 versions · all retained · 1 fresh train

sage_baseline

fresh train

date28 Sept

auroc0.911

fine-tuned from v1

delta+4.2%

date14 Oct

auroc0.918

fine-tuned from v2

delta+3.1%

date4 Nov

auroc0.927

fine-tuned from v2

delta+2.0%

date8 Nov

auroc0.926

fine-tuned from v3

delta+1.5%

date18 Nov

auroc0.940

fine-tuned from v4

delta+1.8%

date30 Nov

auroc0.934

fine-tuned from v4

delta+1.0%

date2 Dec

auroc0.929

active

fine-tuned from v5

delta+0.8%

date5 Dec

auroc0.946

FAQ

Frequently asked questions

What kind of models does Langsat train?

Langsat trains the right model for your data's shape. For a single table it trains tabular models such as MLP, FT-Transformer, ResNet, or TabNet; for two or more linked tables it builds a graph and trains a graph neural network (SAGE, GATv2, or GIN) on the joined structure. It auto-detects keys and relationships and picks a strong default — or you can choose.

Can I export a trained model?

No. A trained model can't be downloaded or exported on any plan, including Team — it isn't provided as a file you can take with you. You use your model through the prediction API and dashboards, and you can export chart and query results to CSV at any time.

Read the Terms

How do I get predictions out of a trained model?

Through a REST prediction API and in-app dashboards. Once a model is trained you can call the API to score new records — one at a time or in batches — pick the prediction mode that fits your question, or explore predictions visually. There's no separate deployment step on your side.

Can I update a model on new data without retraining from scratch?

Yes — fine-tune it. A warm fine-tune adapts your existing model to new data in minutes and keeps what it already learned; a cold fine-tune rebuilds from your full history when the data has shifted. You can freeze the model's foundation for speed and scale, or unfreeze it to adapt more deeply.

How do predictions stay accurate over time?

Langsat watches the data flowing into your models for drift, flags when things shift, and lets you fine-tune or retrain on fresh data — so your predictions stay reliable long after launch instead of going stale.

Train your first model

Start free Talk to sales

Langsat Data Science Predict anything from any data

Every Data Analytics feature, plus the model layer.

One table or many — Langsat takes either.

A single CSV or table.

Two or more tables, linked.

Four modes, picked for the question you're asking.

Let Langsat pick the best approach.

Pure data analysis — no model in the loop.

Score an entity that was in the training set.

Predict for entities never seen before.

Ten models in the registry. Pick one — or race all of them.

Don't pick a model — race them all.

Supervised or unsupervised — six task families, every common subtask.

Regression

Binary Classification

Multi-class Classification

Clustering

Anomaly Detection

Link Prediction

Every task gets the metrics that actually fit it.

One endpoint. Any mode.

INPUT MONITORING

Input Monitoring

FINETUNE › KEEP THE MODEL HONEST

Drop in the new data.

Schema check + delta count.

Resume from the last checkpoint.

Fine-tune lineage 8 versions · all retained · 1 fresh train

Frequently asked questions

Train your first model

FINETUNE