Agents that build the model, run it, and show you the score.
Purpose-built agents that turn your data into forecasts, scores, and recommendations. They run inside your environment, every step verified, with performance you can see.
churn_risk_v7 · scoring pipelineverified ✓ · in production
sourcecustomers
features32 cols
modelxgb · udxf
scorep_churn
gateapproved
exprcustomers.features(...).predict(churn_risk_v7).filter(p_churn > 0.82)runs onSnowflake · UDXF over Arrow Flight · cache HIT for unchanged partitionscommit@d41f0c2a · analytics team · approved by K. Osei
performance · last 30 days
0.91AUC
−38%false positives
1.2Mscored / day
v7vs v6 · +0.03 AUC
precision @ top decile · v7— baseline
where models stalli / iii
Most models die between the notebook and the decision.
A data scientist proves it works. Then it waits: for a pipeline, for a review, for someone to trust the number enough to act on it.
stall 01Features drift from the definitions the business usesThe model’s "active customer" and the dashboard’s "active customer" quietly diverge. Nobody notices until the numbers disagree in a meeting.
stall 02Nobody can say why the score is what it isA decision gets challenged. The path from raw rows to the prediction lives in three notebooks and someone’s memory.
stall 03Every retrain recomputes everythingWarehouse bills climb. The same features get rebuilt for each experiment, and the pipeline is too expensive to run often.
the decision loopii / iii
One expression from raw data to a decision you can defend.
Features, model, score, and gate are one Xorq expression. It compiles to your warehouse, caches what hasn’t changed, and records every version in the Ledger.
Features from blessed sourcesThe agent composes features on the same Ledger definitions the business reports from. No parallel truth.
Model runs where the data isTraining and inference execute as UDXFs inside the expression, streaming over Arrow Flight. Bulk rows never leave the warehouse.
Scores carry their lineageEvery prediction traces to model version, feature set, and source rows. Performance is measured on the same path, not in a side notebook.
A person promotes to productionPromotion is a commit. Rollback is a commit. The gate is human, and the Ledger remembers who opened it.
ledger · churn_riskversions
v7AUC 0.91 · 32 features · xgbproduction
v6AUC 0.88 · 29 features · xgbrolled back
v5AUC 0.86 · 24 features · logregarchived
lineageraw.events → stg.customers → features_v7 → churn_risk_v7cachefeatures replayed for 27/30 partitions · 3 recomputedpromotedK. Osei · @d41f0c2a · 4 days ago
architecture · decision agentsiii / iii
The model is a step in the query, not a system beside it.
Xorq runs the whole loop inside your perimeter. Your warehouse does the heavy lifting; Xorq does the bookkeeping.
Copy our address and paste it into whichever email client you use.
consent — cookies.txt
$
We use cookieless analytics to measure traffic. With your consent, we also enable product
analytics, session recording, and Google Ads conversion tracking, which store data in
your browser. Read our
Privacy Policy.
