Private formulation intelligence for coatings R&D.
A visual demo of how a client-owned ACRAZO LLM can combine public chemistry context with private formulas, TDS/MSDS files, old experiments, customer complaints, and technical service history — inside the customer’s own environment.
ACRAZO Science Formulation AI
Goal: improve adhesion + corrosion resistance without increasing cure time or VOC.
The buyer sees a complete R&D workflow, not a generic chat page.
The demo is intentionally designed as a guided workflow. It is powerful enough for a CEO, CTO, or R&D director to understand the value, but realistic enough to build as a private pilot with templates, retrieval, structured outputs, and LLM reasoning.
Public starting context
Coatings chemistry, substrate types, binder families, common failure modes, public TDS logic, and technical vocabulary.
Private upload
TDS, MSDS, old formulations, test results, customer complaints, sales notes, and technical service history are uploaded only inside the client environment.
Question engine
The system asks missing questions: substrate preparation, humidity, film thickness, curing window, pigment volume, pretreatment, and field conditions.
Experiment narrowing
Instead of proposing thousands of random trials, the workflow narrows virtual combinations into a short, defensible DOE plan.
Decision output
Five prioritized experiments, predicted tradeoffs, customer-ready response, and internal technical report draft.
The visual is a demo metaphor: ingredients, failure modes, test history, customer requirements, and target properties are connected inside the client-controlled environment.
Demo input: field failure + formulation target
User note: “Customer reports edge corrosion and adhesion loss after 480h salt spray on phosphated steel. Current 2K epoxy primer passes initial adhesion but fails after humidity cycle. Need stronger corrosion resistance, same viscosity window, no major cost increase.”
Example AI output shown on the demo page.
The output looks serious without requiring impossible automation on day one. In the pilot, these cards can be generated from the customer’s private documents and controlled templates.
Likely failure causes
Insufficient barrier at edges, pretreatment inconsistency, pigment/binder imbalance after humidity cycling, possible under-cure window, and substrate contamination risk.
Formulation direction
Compare zinc phosphate / aluminum barrier pigment balance, silane pretreatment support, crosslink density adjustment, and lower water uptake resin package.
Customer answer
Generate a technical service response explaining likely causes, proposed corrective tests, and a controlled improvement plan without revealing internal formula details.
| Trial | Purpose | Change to test | Expected signal | Decision rule |
|---|---|---|---|---|
| E-01 | Edge corrosion barrier | Increase lamellar barrier package within viscosity limit | Reduced creep at scribe | Keep if salt spray improves without adhesion loss |
| E-02 | Substrate interface | Add silane-assisted pretreatment / primer interface check | Better wet adhesion after humidity cycle | Keep if cross-hatch remains stable |
| E-03 | Water uptake | Adjust resin/crosslink density while maintaining pot life | Lower blistering tendency | Keep if dry time remains acceptable |
| E-04 | Anti-corrosion pigment | Compare phosphate/chromate-free inhibitor balance | Better corrosion delay | Keep if cost impact stays within range |
| E-05 | Field condition robustness | Test film thickness and cure window sensitivity | Reduced failure variability | Keep if performance holds under low-temperature cure |
What the pilot proves in 60 days.
- Can the private LLM answer technical questions using the customer’s own documents?
- Can it convert scattered history into a structured DOE plan?
- Can it generate a customer-ready technical service response?
- Can R&D, sales, and technical service use one controlled knowledge layer?
What it does not overpromise.
This demo does not claim automatic discovery of a perfect formula from nothing. It shows a realistic private AI decision workflow: retrieval from private documents, structured reasoning, expert review, experiment ranking, and reusable technical output templates.