Hans-Made Research · Codex Resonance

Predictive Sensory
Evaluation.

Read a molecule's taste, shelf-life, safety, and bioavailability from its structure — before the bench. Deterministic, zero training data, milliseconds per compound. A pre-bench prioritizer that tells your panel and your assays where to look first.

Correlated against a third-party University of Manitoba e-tongue (Richardson Centre) · R² 0.95.

02
The problem

The expensive work runs on everything — because you can't tell in advance what matters.

To find which compound makes an isolate taste bitter, you fractionate it, taste it on a trained panel, and confirm by LC-MS / NMR — over weeks, with the physical sample. And taste is only the first question: will it stay good, stay safe, and absorb? Each is another assay, another lab, another sample.

So the panel, the instruments, and the material are spent on the whole list — not the few compounds that drive the result.

03
What it is

One input. One engine. A tiered dossier out.

Input
A structureA SMILES, or a CSV / TSV / JSON panel. No physical sample, no instrument.
Engine
The readDeterministic, structure-only physics. Zero training data. Milliseconds per compound, on a laptop.
Output
The dossierTaste · product-readiness · full safety & bioavailability — with the mechanism named on every call.

The same structure always returns the same answer. Nothing is trained, so it reads molecules that have never been made — exactly where there is no data to learn from.

04
Where it fits

A layer in front of your pipeline — not a replacement for it.

Your candidates
The listThe compounds in an isolate, hydrolysate, or candidate set.
Codex read
SecondsFlags the culprits, the liabilities, and the masker to try — structure-only.
Focused bench
Only what's flaggedSensory panel · safety check · shelf-life — run on the handful that matter.
Decision
FasterFewer iterations, fewer sessions, less wasted material.

It does not replace your panel or your assays. It tells them where to look first.

05
Product features

The specification.

UsePredict taste & off-notes, aroma, shelf-life, safety liabilities, and bioavailability — from structure alone.
InputA SMILES, or a CSV / TSV / JSON panel. No physical sample required.
Throughput~17 ms for the taste read; a full multi-layer dossier in well under a second per compound. A panel in seconds, on a laptop — no GPU.
Operating principleStructure-only physics; zero training data; deterministic.
ControlSame structure → same answer. Shuffle the labels → the signal collapses to chance.
CoverageTaste (7 modalities), aroma, oxidative & light shelf-life, structural safety alerts (reactive-aldehyde / oxidation), bioavailability, mixtures & masking, genetic perception.
AccuracyR² 0.95 vs a third-party e-tongue; 88% balanced accuracy / MCC 0.76 on 500 blind compounds; ~89% off-note recall.
ExplainabilityEvery call names the structural mechanism that fired — not a black box.
OutputA per-compound dossier, TXT + JSON, tiered by subscription.
06
Why the score is real

Four controls for an engine with no training set to overfit.

1 · Out-of-distribution
75–94%Holds up on sources it never saw — non-Western flavour traditions, held-out databases — where chance is 50%.
2 · Noise stress
Graceful, not a cliffErode the structure and the call degrades smoothly. An overfit model crashes to chance the moment the input is imperfect.
3 · Explainability
Mechanism on every callTAS2R ionic binding, polyphenol bitterness, sugar-glycoside sweetness — the same logic a flavour chemist uses.
4 · Real-world
R² 0.95Third-party University of Manitoba e-tongue, measured vs predicted-from-structure. Shuffle the labels → collapses to chance.

Determinism + zero training + the shuffle control = physics, not a fit. We also name where it's weak — poorly-soluble compounds, fine potency within one series — which is why the wins are believable.

07
Value proposition

Focus the spend. Move faster.

08
A worked example

One panel in. The culprits and the liabilities out.

Four compounds from a plant-protein isolate, read from structure — Tier 3, before any sample was made:

CompoundTasteAroma / shelf-lifeSafety
Sinapine (canola)Bitter ✗musty; photolabilestructural safety alert — moderate
Decadienal (oxidation)coolingsharp; fast oxidationreactive aldehyde (Michael/Schiff) — top severity
Genistein (soy)Bitter ✗photolabilelow
Sucrose (control)Sweet ✓stableclean

Two bitter culprits to chase, a structural safety alert and a reactive aldehyde flagged, two compounds that need light-protective packaging — and the benign one cleared.

The bench then confirms only what's flagged. Across 500 blind compounds the off-note call holds at 88% balanced accuracy — and the misses are always shown, never hidden.

09
The market

One read, in front of a five-figure characterization spend.

$35Bplant-protein market by 2030 (7.9% CAGR)
$1.3–2.2Bbitter-blocker & flavour-masking market
$8–20kto characterize one compound across 5–7 labs, over weeks
>$1Mto run a 100-compound program on everything

The engine reads all of them in one pass — so the five-figure assays run only on the handful each layer flags. The same read that prioritizes a food isolate prioritizes a crop-protection candidate too: one engine across the food-and-field value chain, the adjacent uses open as the proof compounds accumulate.

A layer in front of the spend — not a replacement for it.

10
How to engage

Pick the read you need. Start with a blind pilot.

You sendYou get backTier
name, smilestaste call + off-note culprits + ranked masker1 · Taste
+ storage, solubility+ aroma + shelf-life (oxidative / light)2 · Product-readiness
+ dose, genotype+ full safety stack + bioavailability + genetic3 · Full dossier

Send your isolate, hydrolysate, or candidate list as a CSV or JSON. We read each compound from structure and return the taste call, the mechanism, the ranked masker, and the score. You check the results against your own panel and lab.

Dustin Hansley · Hans-Made Research Inc. · Winnipeg, Manitoba
dustinhansmade@gmail.com · 204-333-0234

← Back to the platform