+
1 · Structure → numbers
Every molecule becomes a 2,048-bit Morgan fingerprint (which substructures it contains) plus a block of physicochemical descriptors (logP, MW, TPSA, H-bonding…). That one shared vector feeds every model — the same structure-to-property representation QSAR has leaned on for decades.
+
2 · Taste — 6 trained heads + 2 rules
Six random-forest classifiers (sweet, bitter, umami…) each return a probability, scored by an honest held-out CV-AUROC. Sour and salty are solution / ionic effects, not molecule-shape ones, so they're transparent rules — not faked models. A sweetness-intensity regressor estimates relative-to-sucrose potency.
+
3 · Aroma — 24 odor heads
One random forest per descriptor (citrus, floral, woody…), trained on public-domain odor text. A head ships only if it clears CV-AUROC ≥ 0.70 — 24 survive, each shown with its own score. It reads presence, not intensity (free text carries none) — an honest ceiling, stated in the UI.
+
4 · Honest by design
Every value is tagged measured / predicted / estimate, so nothing reads as more precise than it is. Where a quantitative feature needs data we can't ship free-commercially (odor thresholds, panel intensities), the UI says so — and it lights up with your data.
+
5 · Formulation, not just molecules
The Formulation Studio reads a whole recipe before you pour — weighting each ingredient by odor impact, aggregating the blend's note-profile, flagging the overpowering component, and closing the gap to your target. That's single-molecule ML turned into a bench tool.
+
6 · On-prem & commercial-clean
Nothing leaves the box — a read makes no cloud calls. The shipped models train only on public-domain or permissively-licensed data, so the commercial edition stays clean (provenance tracked in the repo).
+