How our AI works

AI that learns from your batches, and shows its evidence

Oqelvia uses two kinds of AI: prediction models trained on a team's own data, and a language model that interprets analysis results. Rule-based specification checks and standard statistics stay underneath as the verified foundation the AI builds on.

/01

Trained prediction models

Data it uses
A team's own batch records (composition and process values per batch).
What it produces
Learns how inputs drive a quality outcome such as water activity, then predicts new batches with a 95% range.
How you can check it
Every model shows type, version, number of training batches, cross-validated R², RMSE and MAE, and how it compares with a simple average baseline.
/02

AI interpretation (language model)

Data it uses
A completed analysis stored in the workspace: values, specifications and pass/fail results.
What it produces
Writes a plain-language reading of the result: what is out of range, likely causes, and what to check next on the bench.
How you can check it
Each note is labelled with the model used and the time it was generated, and is marked for scientist review.
Published model information(02)

A real model card, computed in front of you

This model is trained in your browser each time the page loads, using the same engine scientists use in the workspace. The numbers are calculated, not written in.

Model card · water activity predictor

RR-1.0 · trained live in your browser just now

Model type
Ridge regression (L2, λ = 1)
Version
RR-1.0
Target
Water activity (aw)
Inputs
Moisture %, protein %, fat %
Training examples
36
Validation
5-fold cross-validation
CV R²
0.973
CV RMSE
0.0079
CV MAE
0.0073
Mean-baseline RMSE
0.0486

Standardized coefficients: moisture +0.0433 · protein -0.0121 · fat +0.0034

Training data: 36 synthetic reference batches generated from a known formula with added noise, so anyone can reproduce this result. It shows how the method is trained and evaluated; it is not a laboratory validation. In a workspace, the same engine trains on a team's own uploaded batch records.

Observed (x) vs out-of-fold predicted (y) aw. Dashed line = perfect prediction.
In the main workflow(03)

From data to an AI-backed decision

  1. 01

    Upload batches

    Import a CSV of production or pilot batches.

  2. 02

    Train

    Pick the outcome to predict and the inputs; the model trains in seconds.

  3. 03

    Check the evidence

    Review cross-validated accuracy against the baseline before trusting it.

  4. 04

    Predict

    Enter a planned formulation and see the predicted value and range.

  5. 05

    Interpret

    Run analyses and get an AI reading of what the results mean.

  6. 06

    Confirm on the bench

    A qualified scientist validates before any release decision.

Train a model on your data